Hi! If you like this piece and want to support my independent reporting and analysis, why not subscribe to my premium newsletter? It’s $70 a year, or $7 a month, and in return you get a weekly newsletter that’s usually anywhere from 5000 to 18,000 words, including vast, detailed analyses of NVIDIA, Anthropic and OpenAI’s finances, and the AI bubble writ large. I just put out a massive Hater’s Guide To The SaaSpocalypse, as well as the Hater’s Guide to Adobe. It helps support free newsletters like these!

The entire AI bubble is built on a vague sense of inevitability — that if everybody just believes hard enough that none of this can ever, ever go wrong that at some point all of the very obvious problems will just go away.

Sadly, one cannot beat physics.

Last week, economist Paul Kedrosky put out an excellent piece centered around a chart that showed new data center capacity additions (as in additions to the pipeline, not brought online) halved in the fourth quarter of 2025 (per data from Wood Mackenzie):

  Wood Mackenzie’s report framed it in harsh terms:

US data-centre capacity additions halved from Q3 to Q4 2025 as load-queue challenges persisted. The decline underscores the difficulties of the current development environment and signals a resulting focus on existing pipeline projects. While Texas extended its pipeline capacity lead in Q4 2025, New Mexico, Indiana and Wyoming saw greater relative growth. Planned capacity continues to be weighted by new developers with a small number of massive, speculative projects, targeting in particular the South and Southwest. New Mexico owes its growth to a single, massive, speculative project by New Era Energy & Digital in Lea County. 

As I said above, this refers only to capacity that’s been announced rather than stuff that’s actually been brought online, and Kedrosky missed arguably the craziest chart — that of the 241GW of disclosed data center capacity, only 33% of it is actually under active development:

The report also adds that the majority of committed power (58%) is for “wires-only utilities,” which means the utility provider is only responsible for getting power to the facility, not generating the power itself, which is a big problem when you’re building entire campuses made up of power-hungry AI servers. 

WoodMac also adds that PJM, one of the largest utility providers in America, “...remains in trouble, with utility large load commitments three times as large as the accredited capacity in PJM’s risked generation queue,” which is a complex way of saying “it doesn’t have enough power.” 

This means that fifty eight god damn percent of data centers need to work out their own power somehow. WoodMac also adds there is around $948 billion in capex being spent in totality on US-based data centers, but capex growth decelerated for the first time since 2023. Kedrosky adds:

The total announced pipeline looks huge at 241 GW — about twice US peak electricity demand — but most of it is not real. Only a third is under construction, with the rest a mix of hopeful permits, speculative land deals, and projects that assume power sources nobody has actually built yet. In particular, much of it assumes on-site gas plants, a fraught assumption given current geopolitics.

The most serious problem is in the mid-Atlantic. Regional grid operator PJM has made power commitments to data centers at roughly three times the rate that new generation is actually coming online. Someone is going to be waiting a very long time, or paying a lot more than they expected, or both.

Let’s simplify:

• Only 33% of announced US data centers are actually being built, with the rest in vague levels of “planning.” That’s about 79.53GW of power, or 61GW of IT load.
  • “Active development” also refers to anything that is (and I quote) “...under development or construction,” meaning “we’ve got the land and we’re still working out what to do with it.
  • This is pretty obvious when you do the maths. 61GW of IT load would be hundreds of thousands of NVIDIA GB200 NVL72 racks — over a trillion dollars of GPUs at $3 million per 72-GPU rack — and based on the fact there were only $178.5 billion in data center debt deals last year, I don’t think many of these are actually being built right now.
• Even if they were, there’s not enough power for them to turn on.
• NVIDIA claims it will sell $1 trillion of GPUs between 2025 and 2027, and as I calculated previously, it sells about 1.6GW (in IT load terms, as in how much power just the GPUs draw) of GPUs every quarter, which would require at least 1.95GW of power just to run, when you include all the associated gear and the challenges of physically getting power.
• None of this data talks about data centers actually coming online.

How Much Actual Data Center Capacity Came Online In 2025? My Estimate: 3GW of IT Load  

The term you’re looking for there is data center absorption, which is (to quote Data Center Dynamics) “...the net growth in occupied, revenue-producing IT load,” which grew in America’s primary markets from 1.8GW in new capacity in 2024 to 2.5GW of new capacity in 2025 according to CBRE.  

Definition sidenote! “Colocation” space refers to data center space built that is then rented out to somebody else, versus data centers explicitly built for a company (such as Microsoft’s Fairwater data centers). What’s interesting is that it appears that some — such as Avison Young — count Crusoe’s developments (such as Stargate Abilene) as colocation construction, which makes the collocation numbers I’ll get to shortly much more indicative of the greater picture.

The problem is, this number doesn’t actually express newly-turned-on data centers. Somebody expanding a project to take on another 50MW still counts as “new absorption.” 

Things get more confusing when you add in other reports. Avison Young’s reports about data center absorption found 700MW of new capacity in Q1 2025, 1.173GW in Q2, a little over 1.5GW in Q3 and 2.033GW in Q4 (I cannot find its Q3 report anywhere), for a total of 5.44GW, entirely in “colocation,” meaning buildings built to be leased to others.

Yet there’s another problem with that methodology: these are facilities that have been “delivered” or have a “committed tenant.” “Delivered” could mean “the facility has been turned over to the client, but it’s literally a powered shell (a warehouse) waiting for installation,” or it could mean “the client is up and running.” A “committed tenant” could mean anything from “we’ve signed a contract and we’re raising funds” (such as is the case with Nebius raising money off of a Meta contract to build data centers at some point in the future).

We can get a little closer by using the definitions from DataCenterHawk (from whichAvison Young gets its data), which defines absorption as follows: 

To measure demand, we want to know how much capacity was leased up by customers over a specific period of time. At datacenterHawk we calculate this quarterly. The resulting number is what’s called absorption.

Let’s say DC#1 has 10 MW commissioned. 9 MW are currently leased and 1 MW is available. Over the course of a quarter, DC#1 leases up that last MW to a few tenants. Their absorption for the quarter would be 1 MW. It can get a little more complicated but that’s the basic concept.

That’s great! Except Avison Young has chosen to define absorption in an entirely different way — that a data center (in whatever state of construction it’s in) has been leased, or “delivered,” which means “a fully ready-to-go data center” or “an empty warehouse with power in it.” 

CBRE, on the other hand, defines absorption as “net growth in occupied, revenue-producing IT load,” and is inclusive of hyperscaler data centers. Its report also includes smaller markets like Charlotte, Seattle and Minneapolis, adding a further 216MW in absorption of actual new, existing, revenue-generating capacity.

So that’s about 2.716GW of actual, new data centers brought online. It doesn’t include areas like Southern Virginia or Columbus, Ohio — two massive hotspots from Avison Young’s report — and I cannot find a single bit of actual evidence of significant revenue-generating, turned-on, real data center capacity being stood up at scale. DataCenterMap shows 134 data centers in Columbus, but as of August 2025, the Columbus area had around 506MW in total according to the Columbus Dispatch, though Cushman and Wakefield claimed in February 2026 that it had 1.8GW.

Things get even more confusing when you read that Cushman and Wakefield estimates that around 4GW of new colocation supply was “delivered” in 2025, a term it does not define in its actual report, and for whatever reason lacks absorption numbers. Its H1 2025 report, however, includes absorption numbers that add up to around 1.95GW of capacity…without defining absorption, leaving us in exactly the same problem we have with Avison Young. 

Nevertheless, based on these data points, I’m comfortable estimating that North American data center absorption — as the IT load of data centers actually turned on and in operation — was at around 3GW for 2025, which would work out to about 3.9GW of total power.

And that number is a fucking disaster.

It Is Currently Taking 6 Months To Install A Quarter of NVIDIA’s GPU Sales, Calling Into Question The Logic of Buying More GPUs 

Earlier in the year, TD Cowen’s Jerome Darling told me that GPUs and their associated hardware cost about $30 million a megawatt. 3GW of IT load (as in the GPUs and their associated gear’s power draw) works out to around $90 billion of NVIDIA GPUs and the associated hardware, which would be covered under NVIDIA’s “data center” revenue segment:

America makes up about 69.2% of NVIDIA’s revenue, or around $149.6 billion in FY2026 (which runs, annoyingly, from February 2025 to January 2026). NVIDIA’s overall data center segment revenue was $195.7 billion, which puts America’s data center purchases at around $135 billion, leaving around $44 billion of GPUs and associated technology uninstalled.

With the acceleration of NVIDIA’s GPU sales, it now takes about 6 months to install and operationalize a single quarter’s worth of sales. Because these are Blackwell (and I imagine some of the new next generation Vera Rubin) GPUs, they are more than likely going to new builds thanks to their greater power and cooling requirements, and while some could in theory be going to old builds retrofitted to fit them, NVIDIA’s increasingly-centralized (as in focused on a few very large customers) revenue heavily suggests the presence of large resellers like Dell or Supermicro (which I’ll get to in a bit) or the Taiwanese ODMs like Foxconn and Quanta who manufacture massive amounts of servers for hyperscaler buildouts. 

I should also add that it’s commonplace for hyperscalers to buy the GPUs for their colocation partners to install, which is why Nebius and Nscale and other partners never raise more than a few billion dollars to cover construction costs. 

It’s becoming very obvious that data center construction is dramatically slower than NVIDIA’s GPU sales, which continue to accelerate dramatically every single quarter.

Even if you think AI is the biggest most hugest and most special boy: what’s the fucking point of buying these things two to four years in advance? Jensen Huang is announcing a new GPU every year! 

By the time they actually get all the Blackwells in Vera Rubin will be two years old! And by the time we install those Vera Rubins, some other new GPU will be beating it! 

There Is Only 5GW of Global Data Center Capacity Actually Under Construction, And Every Huge, Multi-Gigawatt Project You Read Is Going To Take 2 to 4 Years Or More To Complete — And Wood Mackenzie Believes Capex Growth Will Slow In 2026

Before we go any further, I want to be clear how difficult it is to answer the question “how long does a data center take to build?”. You can’t really say “[time] per megawatt” because things become ever-more complicated with every 100MW or so. As I’ll get into, it’s taken Stargate Abilene two years to hit 200MW of power.

Not IT load. Power. 

Anyway, the question of “how much data center capacity came online?” is pretty annoying too. 

Sightline’s research — which estimated that “almost 6GW of [global data center power] capacity came online last year” — found that while 16GW of capacity was slated to come online in 2026 across 140 projects, only 5GW is currently under construction, and somehow doesn’t say that “maybe everybody is lying about timelines.”

Sightline believes that half of 2026’s supposed data center pipeline may never materialize, with 11GW of capacity in the “announced” stage with “...no visible construction progress despite typical build timelines of 12-18 months.” “Under construction” also can mean anything from “a single steel beam” to “nearly finished.”

These numbers also are based on 5GW of capacity, meaning about 3.84GW of IT load, or about $111.5 billion in GPUs and associated gear, or roughly 57.5% of NVIDIA’s FY2026 revenue that’s actually getting built.

Sightline (and basically everyone else) argues that there’s a power bottleneck holding back data center development, and Camus explains that the biggest problem is a lack of transmission capacity (the amount of power that can be moved) and power generation (creating the power itself): 

The biggest driver of delay is simple: our power system doesn’t have enough extra transmission capacity and generation to serve dozens of gigawatts of new, high-utilization demand 100% of the time. Data centers require round-the-clock power at levels that rival or exceed the needs of small cities, and building new transmission infrastructure and generation requires years of permitting, land acquisition, supply chain management, and construction.

Camus adds that America also isn’t really prepared to add this much power at once:

Inside utilities, planners and engineers are working diligently to connect new loads. But the tools available to planners were built for extending power lines to new neighborhoods or upgrading equipment as communities grow. They weren’t designed to analyze 50 new service requests of 100 MW each, all while new generation applications pile up.

As a result, planners and engineers are overwhelmed; they’re stuck working to review new applications while simultaneously configuring new tools that are better equipped for the scale of this challenge. And unlike generation interconnection, which has well-defined steps across most ISOs and utilities, the process for evaluating large loads is often much more ad hoc. This makes adopting the right tools much more difficult too. In fact, the majority of utilities and ISO/RTOs are still developing formal study procedures.

Nevertheless, I also think there’s another more-obvious reason: it takes way longer to build a data center than anybody is letting on, as evidenced by the fact that we only added 3GW or so of actual capacity in America in 2025. NVIDIA is selling GPUs years into the future, and its ability to grow, or even just maintain its current revenues, depends wholly on its ability to convince people that this is somehow rational.

Let me give you an example. OpenAI and Oracle’s Stargate Abilene data center project was first announced in July 2024 as a 200MW data center. In October 2024, the joint venture between Crusoe, Blue Owl and Primary Digital Infrastructure raised $3.4 billion, with the 200MW of capacity due to be delivered “in 2025.” A mid-2025 presentation from land developer Lancium said it would have “1.2GW online by YE2025.” In a May 2025 announcement, Crusoe, Blue Owl, and Primary Digital Infrastructure announced the creation of a $15 billion joint vehicle, and said that Abilene would now be 8 buildings, with the first two buildings being energized by the “first half of 2025,” and that the rest would be “energized by mid-2026.” Each building would have 50,000 GPUs, and the total IT load is meant to be 880MW or so, with a total power draw of 1.2GW. 

I’m not interested in discussing OpenAI not taking the supposedly-planned extensions to Abilene because it never existed and was never going to happen. 

In December 2025, Oracle stated that it had “delivered” 96,000 GPUs, and in February, Oracle was still only referring to two buildings, likely because that’s all that’s been finished. My sources in Abilene tell me that Building Three is nearly done, but…this thing is meant to be turned on in mid-2026. Developer Mortensen claims the entire project will be completed by October 2026, which it obviously, blatantly won’t.

The AI Industry Is Trying To Hide The Incredibly Slow Growth of Data Centers, And Is Using The Media To Do So

I hate to speak in conspiratorial terms, but this feels like a blatant coverup with the active participation of the press. CNBC reported in September 2025 that “the first data center in $500 billion Stargate project is open in Texas,” referring to a data center with an eighth of its IT load operational as “online” and “up and running,” with Crusoe adding two weeks later that it was “live,” “up and running” and “continuing to progress rapidly,” all so that readers and viewers would think “wow, Stargate Abilene is up and running” despite it being months if not years behind schedule.

At its current rate of construction, Stargate Abilene will be fully built sometime in late 2027. Oracle’s Port Washington Data Center, as of March 6 2026, consisted of a single steel beam. Stargate Shackelford Texas broke ground on December 15 2025, and as of December 2025, construction barely appears to have begun in Stargate New Mexico. Meta’s 1GW data center campus in Indiana only started construction in February 2026. 

And, despite Microsoft trying to mislead everybody that its Wisconsin data center had ‘arrived” and “been built,” looking even an inch deeper suggests very little has actually come online” — and, considering the first data center was $3.3 billion (remember: $14 million a megawatt just for construction), I imagine Microsoft has successfully brought online about 235MW of power for Fairwater.

What Microsoft wants you to think is it brought online gigawatts of power (always referred to in the future tense), because Microsoft, like everybody else, is building data centers at a glacial pace, because construction takes forever, even if you have the power, which nobody does!

The concept of a hundred-megawatt data center is barely a few years old, and I cannot actually find a built, in-service gigawatt data center of any kind, just vague promises about theoretical Stargate campuses built for OpenAI, a company that cannot afford to pay its bills. 

Everybody keeps yammering on about “what if data centers don’t have power” when they should be thinking about whether data centers are actually getting built. Microsoft proudly boasted in September 2025 about its intent to build “the UK’s largest supercomputer” in Loughton, England with Nscale, and as of March 2026, it’s literally a scaffolding yard full of pylons and scrap metal. Stargate Abilene has been stuck at two buildings for upwards of six months. 

Here’s what’s actually happening: data center deals are being funded by eager private credit gargoyles that don’t know shit about fuck. These deals are announced, usually by overly-eager reporters that don’t bother to check whether the previous data centers ever got built, as massive “multi-gigawatt deals,” and then nobody follows up to check whether anything actually happened. 

All that anybody needs to fund one of these projects is an eager-enough financier and a connection to NVIDIA. All Nebius had to do to raise $3.75 billion in debt was to sign a deal with Meta for data center capacity that doesn’t exist and will likely take three to four years to build (it’s never happening). Nebius has yet to finish its Vineland, New Jersey data center for Microsoft, which was meant to be “at 100MW” by the end of 2025, but appears to have only had 50MW (the first phase) available as of February 2026. 

I’m just gonna come out and say it: I think a lot of these data center deals are trash, will never get built, and thus will never get paid. The tech industry has taken advantage of an understandable lack of knowledge about construction or power timelines in the media to pump out endless stories about “data center capacity in progress” as a means of obfuscating an ever-growing scandal: that hundreds of billions of NVIDIA GPUs got sold to go in projects that may never be built.

These things aren’t getting built, or if they’re getting built, it’s taking way, way longer than expected, which means that interest on that debt is piling up. The longer it takes, the less rational it becomes to buy further NVIDIA GPUs — after all, if data centers are taking anywhere from 18 months to three years to build, why would you be buying more of them? Where are you going to put them, Jensen?

This also seriously brings into question the appetite that private credit and other financiers have for funding these projects, because much of the economic potential comes from the idea that these projects get built and have stable tenants. Furthermore, if the supply of AI compute is a bottleneck, this suggests that when (or if) that bottleneck is ever cleared, there will suddenly be a massive supply glut, lowering the overall value of the data centers in progress…which are, by the way, all filled with Blackwell GPUs, which will be two or three-years-old by the time the data centers are finally turned on.

That’s before you get to the fact that the ruinous debt behind AI data centers makes them all remarkably unprofitable, or that their customers are AI startups that lose hundreds of millions or billions of dollars a year, or that NVIDIA is the largest company on the stock market, and said valuation is a result of a data center construction boom that appears to be decelerating and even if it wasn’t operating at a glacial pace compared to NVIDIA’s sales.

Not to sound unprofessional or nothing, but what the fuck is going on? We have 241GW of “planned” capacity in America, of which only 79.5GW of which is “under active development,” but when you dig deeper, only 5GW of capacity is actually under construction? 

The entire AI bubble is a god damn mirage. Every single “multi-gigawatt” data center you hear about is a pipedream, little more than a few contracts and some guys with their hands on their hips saying “brother we’re gonna be so fuckin’ rich!” as they siphon money from private credit — and, by extension, you, because where does private credit get its capital from? That’s right. A lot comes from pension funds and insurance companies.

Here’s the reality: data centers take forever. Every hyperscaler and neocloud talking about “contracted compute” or “planned capacity” may as well be telling you about their planned dinners with The Grinch and Godot. The insanity of the AI buildout will be seen as one of the largest wastes of capital of all time (to paraphrase JustDario), and I anticipate that the majority of the data center deals you’re reading about simply never get built.

The fact that there’s so much data about data center construction and so little data about completed construction suggests that those preparing the reports are in on the con. I give credit to CBRE, Sightline and Wood Mackenzie for having the courage to even lightly push back on the narrative, even if they do so by obfuscating terms like “capacity” or “power” in ways that reporters and other analysts are sure to misinterpret.

Hundreds of billions of dollars have been sunk into buying GPUs, in some cases years in advance, to put into data centers that are being built at a rate that means that NVIDIA’s 2025 and 2026 revenues will take until 2028 to 2029 to actually operationalize, and that’s making the big assumption that any of it actually gets built.

I think it’s also fair to ask where the money is actually going. 2025’s $178.5 billion in US-based data center deals doesn’t appear to be resulting in any immediate (or even future) benefit to anybody involved.

I also wonder whether the demand actually exists to make any of this worthwhile, or what people are actually paying for this compute. 

If we assume 3GW of IT load capacity was brought online in America, that should (theoretically) mean tens of billions of dollars of revenue thanks to the “insatiable demand for AI” — except nobody appears to be showing massive amounts of revenue from these data centers. 

Applied Digital only had $144 million in revenue in FY2025 (and lost $231 million making it). CoreWeave, which claimed to have “850MW of active power (or around 653MW of IT load)” at the end of 2025 (up from 420MW in Q1 FY2025, or 323MW of IT load), made $5.13 billion of revenue (and lost $1.2 billion before tax) in FY2025. 

Nebius? $228 million, for a loss of $122.9 million on 170MW of active power (or around 130MW of IT load). Iren lost $155.4 million on $184.7 million last quarter, and that’s with a release of deferred tax liabilities of $182.5 million. Equinix made about $9.2 billion in revenue in its last fiscal year, and while it made a profit, it’s unclear how much of that came from its large and already-existent data center portfolio, though it’s likely a lot considering Equinix is boasting about its “multi-megawatt” data center plans with no discussion of its actual capacity.

And, of course, Google, Amazon, and Microsoft refuse to break out their AI revenues. Based on my reporting from last year, OpenAI spent about $8.67 billion on Azure through September 2025, and Anthropic around $2.66 billion in the same period on Amazon Web Services. As the two largest consumers of AI compute, this heavily suggests that the actual demand for AI services is pretty weak, and mostly taken up by a few companies (or hyperscalers running their own services.) 

At some point reality will set in and spending on NVIDIA GPUs will have to decline. It’s truly insane how much has been invested so many years in the future, and it’s remarkable that nobody else seems this concerned.

Simple questions like “where are the GPUs going?” and “how many actual GPUs have been installed?” are left unanswered as article after article gets written about massive, multi-billion dollar compute deals for data centers that won’t be built before, at this rate, 2030. 

And I’d argue it’s convenient to blame this solely on power issues, when the reality is clearly based on construction timelines that never made any sense to begin with. If it was just a power issue, more data centers would be near or at the finish line, waiting for power to be turned on. Instead, well-known projects like Stargate Abilene are built at a glacial pace as eager reporters claim that a quarter of the buildings being functional nearly a year after they were meant to be turned on is some sort of achievement.

Then there’s the very, very obvious scandal that NVIDIA, the largest company on the stock market, is making hundreds of billions of dollars of revenue on chips that aren’t being installed. It’s fucking strange, and I simply do not understand how it keeps beating and raising expectations every quarter given the fact that the majority of its customers are likely going to be able to use their current purchases in the next decade.

Assuming that Vera Rubin actually ships in 2026, it’s reasonable to believe that people will be installing these things well into 2028, if not further, and that’s assuming everything doesn’t collapse by then. Why would you bother? What’s the point, especially if you’re sitting on a pile of Blackwell GPUs? 

Why are we doing any of this? 

Jensen, How Do All These NVIDIA GPUs Keep Getting To China?

Last week also featured a truly bonkers story about Supermicro, a reseller of GPUs used by CoreWeave and Crusoe, where co-founder Wally Liaw and several other co-conspirators were arrested for selling hundreds of millions of dollars of NVIDIA GPUs to China, with the intent to sell billions more. 

Liaw, one of Supermicro’s co-founders, previously resigned in a 2018 accounting scandal where Supermicro couldn’t file its annual reports, only to be (per Hindenburg Research’s excellent report) rehired in 2021 as a consultant, and restored to the board in 2023, per a filed 8K. 

Mere days before his arrest, Liaw was parading around NVIDIA’s GTC conference, pouring unnamed liquids in ice luges and standing two people away from NVIDIA CEO Jensen Huang. Liaw was also seen congratulating the CEO of Lambda on its new CFO appointment on LinkedIn, as well as shaking hands (along with Supermicro CEO Charles Liang, who has not been arrested or indicted) with Crusoe (the company building OpenAI’s Abilene data center) CEO Chase Lochmiller. 

Supermicro isn’t named in the indictment for reasons I imagine are perfectly normal and not related to keeping the AI party going. Nevertheless, Liaw and his co-conspirators are accused of shipping hundreds of millions of dollars’ worth of NVIDIA GPUs to China through a web of counterparties and brokers, with over $510 million of them shipped between April and mid-May 2025. While the indictment isn’t specific as to the breakdown, it confirms that some Blackwell GPUs made it to China, and I’d wager quite a few.

The mainstream media has already stopped thinking about this story, despite Supermicro being a huge reseller of NVIDIA gear, contributing billions of dollars of revenue, with at least $500 million of that apparently going to China. The fact that Supermicro wasn’t specifically named in the case is enough to erase the entire tale from their minds, along with any wonder about how NVIDIA, and specifically Jensen Huang, didn’t know.

This also isn’t even close to the only time this has happened. Late last year, Bloomberg reported on Singapore-based Megaspeed — a (to quote Bloomberg) “once-obscure spinoff of a Chinese gaming enterprise [that] evolved into the single largest Southeast Asian buyer of NVIDIA chips” — and highlighted odd signs that suggest it might be operating as a front for China. 

As a neocloud, Megaspeed rents out AI compute capacity like CoreWeave, and while NVIDIA (and Megaspeed) both deny any of their GPUs are going to China, Megaspeed, to quote Bloomberg, has “something of a Chinese corporate twin”:

This firm used similar presentation materials to Megaspeed’s, had a nearly identical website to a Megaspeed sub-brand and claimed Megaspeed’s Southeast Asia employees as its own. It’s also posted job ads at and near the Shanghai data center whose rendering was used in Megaspeed’s investor deck — including for engineering work on restricted Nvidia GPUs.

Bloomberg reported that Megaspeed imported goods “worth more than a thousand times its cash balance in 2023,” with two-thirds of its imports being NVIDIA products. The investigation got weirder when Bloomberg tried to track down specific circuit boards that NVIDIA had told the US government were in specific sites:

Data centers aren’t the only Megaspeed facilities Nvidia visited. The vast majority of Megaspeed’s $2.4 billion worth of Bianca boards, the circuit boards that house Nvidia’s top-end GB200 and GB300 semiconductors, were unaccounted for at the sites Nvidia described to Washington. After Bloomberg asked about those products, the chipmaker went to separate Megaspeed warehouses, an Nvidia official said, and confirmed the Bianca boards are there.

This person declined to specify the number observed in storage, nor where and when the chips — imported more than half a year ago — would be put to use. “Building data centers is a complex process that takes many months and involves many suppliers, contractors and approvals,” an Nvidia spokesperson said.

Things get weirder throughout the article, with a Chinese company called “Shanghai Shuoyao” having a near-identical website and investor deck (as mentioned) to Megaspeed, with several of the “computing clusters under construction” actually being in China. 

Things get a lot weirder as Bloomberg digs in, including a woman called “Huang” that may or may not be both the CEO of Megaspeed and an associated company called “Shanghai Hexi,” which is also owned by the Yangtze River Delta project… who was also photographed sitting next to Jensen Huang at an event in Taipei in 2024.

While all of this is extremely weird and suspicious, I must be clear there is no declarative answer as to what’s going on, other than that NVIDIA GPUs are absolutely making it to China, somehow. I also think that it would be really tough for Jensen Huang to not know about it, or for billions of dollars of GPUs to be somewhere without NVIDIA’s knowledge. 

Anyway, Supermicro CEO Charles Liang has yet to comment on Wally Liaw or his alleged co-conspirators, other than a statement from the company that says that their acts were “a contravention of the Company’s policies and compliance controls.” 

Jensen Huang does not appear to have been asked if he knew anything about this — not Megaspeed, not Supermicro, or really any challenging question of any kind for the last few years of his life. 

Huang did, however, say back in May 2025 that there was “no evidence of any AI chip diversion,’ and that the countries in question “monitor themselves very carefully.” 

For legal reasons I am going to speak very carefully: I cannot say that Jensen is wrong, or lying, but I think it’s incredible, remarkable even, that he had no idea that any of this was going on. Really? Hundreds of millions if not billions of dollars of GPUs are making it to China — as reported by The Information in December 2025 — and Jensen Huang had no idea? I find that highly unlikely, though I obviously can’t say for sure.

In the event that NVIDIA had knowledge — which I am not saying it did, of course — this is a huge scandal that, for the most part, nobody has bothered to keep an eye on outside of a few brave souls at The Information and Bloomberg who give a shit about the truth. Has anybody bothered to ask Jensen about this? People talk to him on camera all the time. 

Sidenote: Earlier today, US Senators Jim Banks and Elizabeth Warren issued a letter to Howard Lutnick, Trump' s Commerce Secretary, demanding the Department of Commerce take “all necessary and appropriate actions” to stop the flow of NVIDIA chips to China, including potentially block exports to countries believed to be intermediaries, like Malaysia, Thailand, Vietnam, and Singapore.

The arrest of Liaw has, it seems, ruffled some feathers in Washington, and I would not be shocked to see Huang sat before a congressional inquiry at some point.

I’ll also add that I am shocked that so many people are just shrugging and moving on from Supermicro, which is a major supplier of two of the major neoclouds (Crusoe and CoreWeave) and one of the minors (Lambda, which they also rents cloud capacity to). The idea that a company had no idea that several percentage points of its revenue were flowing directly to China via one of its co-founders is an utter joke.

I hope we eventually find out the truth. Nevertheless, this kind of underhanded bullshit is a sign of desperation on the part of just about everybody involved.

The End of Software Engineering — Hyperscalers Are Forcing AI On Their Workers, Destroying The Quality Of Their Products, and Crashing Their Services

So, I want to explain something very clearly for you, because it’s important you understand how fucked up shit has become: hyperscalers are forcing everybody in their companies to use AI tools as much as possible, tying compensation and performance use to token burn, and actively encouraging non-technical people to vibe-code features that actually reach production. 

In practice, this means that everybody is being expected to dick around with AI tools all day, with the expectation that you burn massive amounts of tokens and, in the case of designers working in some companies, actively code features without ever knowing a line of code. 

“How do I know the last part? Because a trusted source told me — and I’ll leave it at that”

One might be forgiven for thinking this means that AI has taken a leap in efficacy, but the actual outcomes are a labyrinth of half-functional internal dashboards that measure random user data or convert files, spending hours to save minutes of time at some theoretical point. While non-technical workers aren’t necessarily allowed to ship directly to production, their horrifying pseudo-software, coded without any real understanding of anything, is expected to be “fixed” by actual software engineers who are also expected to do their jobs.

These tools also allow near-incompetent Business Idiot software engineers to do far more damage than they might have in the past. LLM use is relatively-unrestrained (and actively incentivized) in at least one hyperscaler, with just about anybody allowed to spin up their own OpenClaw “AI agent” (read: series of LLMs that allegedly can do stuff with your inbox or Slack for no clear benefit, other than their ability to delete all of your emails). In Meta’s case, this ended up causing a severe security breach:

According to internal Meta communications and an incident report seen by The Information, a major security alert occurred last week after a Meta software engineer used an in-house agent tool, similar to OpenClaw, to analyze a technical question that another Meta employee had posted on an internal discussion forum. After doing the analysis, the AI agent posted a response in the discussion forum to the original question, offering advice on the technical issue, according to internal communications. The agent did so without approval from the employee.

According to The Information, Meta systems storing large amounts of company and user-related data were accessible to engineers who didn’t have permission to see them, and was marked a sec-1 incident, the second highest level of severity on an internal scale that Meta uses to rank security incidents. 

The incident follows multiple problems caused at Amazon by its Kiro and Q LLMs. I quote Business Insider’s Eugene Kim: 

On March 2, customers across Amazon marketplaces saw incorrect delivery times when adding items to their carts. The incident led to nearly 120,000 lost orders and roughly 1.6 million website errors. Amazon's AI tool Q was one of the primary contributors that triggered the event, according to an internal review.

On March 5, another outage caused a 99% drop in orders across Amazon's North American marketplaces, resulting in 6.3 million lost orders, one of the internal documents stated. One key factor was a production change that was deployed without using a formal documentation and approval process called Modeled Change Management.

LLMs Are Destroying Big Tech From The Inside

Despite the furious (and exhausting) marketing campaign around “the power of AI code,” I believe that these events are just the beginning of the true consequences of AI coding tools: the slow destruction of the tech industry’s software stack. 

LLMs allow even the most incompetent dullard to do an impression of a software engineer, by which I mean you can tell it “make me software that does this” or “look at this code and fix it” and said LLM will spend the entire time saying “you got this” and “that’s a great solution.” 

The problem is that while LLMs can write “all” code, that doesn’t mean the code is good, or that somebody can read the code and understand its intention (as these models do not think), or that having a lot of code is a good thing both in the present and in the future of any company built using generative code. 

LLM-based code is often verbose, and rarely aligns with in-house coding guidelines and standards, guaranteeing that it’ll take far longer to chew through, which naturally means that those burdened with reviewing it will either skim-read it or feed it into another LLM to work out what the hell to do.

Worse still, LLM use is also entirely directionless. Why is anybody at Meta using an OpenClaw? What is the actual thing that OpenClaw does, other than burn an absolute fuck-ton of tokens? 

Think about this very, very simply for a second: you have given every engineer in the company the explicit remit to write all their code using LLMs, and incentivized them to do so by making sure their LLM use is tracked. You have now massively increased both the operating costs of the company (through token burn costs) and the volume of code being created. 

To be explicit, allowing an LLM to write all of your code means that you are no longer developing code, nor are you learning how to develop code, nor are you going to become a better software engineer as a result. This means that, across almost every major tech company, software engineers are being incentivized to stop learning how to write software or solve software architecture issues. 

If you are just a person looking at code, you are only as good as the code the model makes, and as Mo Bitar recently discussed, these models are built to galvanize you, glaze you, and tell you that you’re remarkable as you barely glance at globs of overwritten code that, even if it functions, eventually grows to a whole built with no intention or purpose other than what the model generated from your prompt. 

Things only get worse when you add in the fact that hyperscalers like Meta and Amazon love to lay off thousands of people at a time, which makes it even harder to work out why something was built in the way it was built, which is even harder when an LLM that lacks any thoughts or intentions builds it. Entire chunks of multi-trillion dollar market cap companies are being written with these things, prompted by engineers (and non-engineers!) who may or may not be at the company in a month or a year to explain what prompts they used. 

We’re already seeing the consequences! Amazon lost hundreds of thousands of orders! Meta had a major security breach! The foundations of these companies are being rotted away through millions of lines of slop-code that, at best, occasionally gets the nod from somebody who has “software engineer” on their resume, and these people keep being fired too, raising the likelihood that somebody who knows what’s going on or why something is built a certain way will be able to stop something bad from happening. 

Remember: Google, Amazon, Microsoft, and Meta all hold vast troves of personal information, intimate conversations, serious legal documents, financial information, in some cases even social security numbers, and all four of them along with a worrying chunk of the tech industry are actively encouraging their software engineers to stop giving a fuck about software. 

Oh, you’re so much faster with AI code? What does that actually mean? What have you built? Do you understand how it works? Did you look at the code before it shipped, or did you assume that it was fine because it didn’t break? 

This is creating a kind of biblical plague within software engineering — an entire tech industry built on reams of unmanageable and unintentional code pushed by executives and managers that don’t do any real work. LLMs allow the incompetent to feign competence and the unproductive to produce work-adjacent materials borne of a loathing for labor and craftsmanship, and lean into the worst habits of the dullards that rule Silicon Valley.

All the Valley knows is growth, and “more” is regularly conflated with “valuable.” The New York Times’ Kevin Roose — in a shocking attempt at journalism — recently wrote a piece celebrating the competition within Silicon Valley to burn more and more tokens using AI models:

An engineer at OpenAI processed 210 billion “tokens” — enough text to fill Wikipedia 33 times — through the company’s artificial intelligence models over the last week, the most of any employee. At Anthropic, a single user of the company’s A.I. coding system, Claude Code, racked up a bill of more than $150,000 in a month.

And at tech companies like Meta and Shopify, managers have started to factor A.I. use into performance reviews, rewarding workers who make heavy use of A.I. tools and chastening those who don’t.

This is the new reality for coders, some of the first white-collar workers to feel the effects of A.I. as it sweeps through the economy. A.I. was supposed to help tech companies boost productivity and cut costs. But it has also created an expensive new status game, known as “tokenmaxxing,” among A.I.-obsessed workers who are desperate to prove how productive they are.

Roose explains that both Meta and OpenAI have internal leaderboards that show how many tokens you’ve used, with one software engineer in Stockholm spending “more than his salary in tokens,” though Roose adds that his company pays for them.

Roose describes a truly sick culture, one where OpenAI gives awards to those who spend a lot of money on their tokens, adding that he spoke with several tech workers who were spending thousands of dollars a day on tokens “for what amount to bragging rights.” Roose also added one more insane detail: that one person found a loophole in Claude’s $20-a-month using a piece of software made by Figma that allowed them to burn $70,000 in tokens.

Despite all of this burn, Roose struggled to find anybody who was able to explain what they were doing beyond “maintaining large, complex pieces of software using coding agents running in parallel,” but managed to actually find one particularly useful bit of information — that all of this might be performative:

They said, by and large, that A.I. coding tools were making them more productive. But some also framed their use of A.I. as a strategic move — a way to signal, to their colleagues and bosses, that they’re keeping up with the times, as the era of human coding appears to be coming to an end.

I do give Roose one point for wondering if “...any of these tokenmaxxers [were] producing anything good, or whether they [were] merely spinning their wheels churning out useless code in an attempt to look busy.” Good job Kevin. 

That being said, I find this story horrifying, and veering dangerously close to the actions of drug addicts and cult followers. Throughout this story in one of the world’s largest newspapers, Roose fails to find a single “tokenmaxxer” making something that they can actually describe, which has largely been my experience of evaluating anyone who talks nonstop about the power of “agentic coding.” 

These people are sick, and are participating in a vile, poisonous culture based on needless expenses and endless consumption. 

Companies incentivizing the amount of tokens you burn are actively creating a culture that trades excess for productivity, and incentivizing destructive tendencies built around constantly having to find stuff to do rather than do things with intention.  They are guaranteeing that their software will be poorly-written and maintained, all in the pursuit of “doing more AI” for no reason other than that everybody else appears to be doing so.

Anybody who actually works knows that the most productive-seeming people are often also the most-useless, as they’re doing things to seem productive rather than producing anything of note. A great example of this is a recent Business Insider interview with a person who got laid off from Amazon after learning “AI” and “vibe coding,” and how surprised they were that these supposed skills didn’t make them safer from layoffs:

At the time of the October layoffs, there was debate around whether AI was the reason.

The company was encouraging us to use AI at the time, but I don't think it took my job. I wrote descriptions for internal products at Amazon, and when I used AI to help, I'd need to ask it to rewrite its output without fluff words. It didn't sound like how people talk. Despite my ethical qualms, I used AI, but, in my opinion, it was nowhere close to replacing my role. Before I was laid off, I helped build an internal site for Amazon using AI. I hadn't really coded before, but with a colleague's help, I learned how to vibe code with a lot of trial and error.

I thought using AI for this project and showcasing different skills would make me more valuable to the company, but in the end, it didn't keep me from being laid off.

To be clear, this person is a victim. They were pressured by Amazon to take up useless skills and build useless things in an expensive and inefficient way, and ended up losing their job despite taking up tools they didn’t like under duress. 

Sidenote: If you read that sentence and suggest that she should’ve used AI better, you are a mark. You are being conned into an unpaid marketing job for AI companies that actively hate you. 

This person was, at one point, actively part of building an internal Amazon site using AI, and had to “learn to vibe code with a lot of trial and error” and the help of a colleague. Was this a good use of her time? Was this a good use of her colleague’s time?

No! In fact, across all of these goddamn AI coding hype-beast Twitter accounts and endless proclamations about the incredible power of AI agents, I can find very few accounts of something happening other than someone saying “yeah I’m more productive I guess.” 

I am certain that at some point in the near future a major big tech service is going to break in a way that isn’t immediately fixable as a result of thousands of people building software with AI coding tools, a problem compounded by the dual brain drain forces of layoffs and a culture that actively empowers people to look busy rather than actually produce useful things.

What else would you expect? You’re giving people a number that they can increase to seem better at their job, what do you think they’re going to do, try and be efficient? Or use these things as much as humanly possible, even if there really isn’t a reason to?

I haven’t even gotten to how expensive all of this must be, in part because it’s hard to fully comprehend. 

But what I do know is that big tech is setting itself up for crisis after crisis, especially when Anthropic and OpenAI stop subsidizing their models to the tune of allowing people to spend $2500 or more on a $200-a-month subscription. 

What happens to the people who are dependent on these models? What happens to the people who forgot how to do their jobs because they decided to let AI write all of their code? Will they even be able to do their jobs anymore?  

Large Language Models are creating Silicon Valley Habsburgs — workers that are intellectually trapped at whatever point they started leaning on these models that were subsidized to the point that their bosses encouraged them to use them as much as humanly possible. While they might be able to claw their way back into the workforce, a software engineer that’s only really used LLMs for anything longer than a few months will have to relearn the basic habits of their job, and find that their skills were limited to whatever the last training run for whatever model they last used was. 

I’m sure there are software engineers using these models ethically, who read all the code, who have complete industry over it and use it as a means of handling very specific units of work that they have complete industry over.

I’m also sure that there are some that are just asking it to do stuff, glancing at the code and shipping it. It’s impossible to measure how many of each camp there are, but hearing Spotify’s CEO say that its top developers are basically not writing code anymore makes me deeply worried, because this shit isn’t replacing software engineering at all — it’s mindlessly removing friction and putting the burden of “good” or “right” on a user that it’s intentionally gassing up.

Ultimately, this entire era is a test of a person’s ability to understand and appreciate friction. 

Friction can be a very good thing. When I don’t understand something, I make an effort to do so, and the moment it clicks is magical. In the last three years I’ve had to teach myself a great deal about finance, accountancy, and the greater technology industry, and there have been so many moments where I’ve walked away from the page frustrated, stewed in self-doubt that I’d never understand something.

I also have the luxury of time, and sadly, many software engineers face increasingly-deranged deadlines set by bosses that don’t understand a single fucking thing, let alone what LLMs are capable of or what responsible software engineering is. The push from above to use these models because they can “write code faster than a human” is a disastrous conflation of “fast” and “good,” all because of flimsy myths peddled by venture capitalists and the media about “LLMs being able to write all code.”

Generative code is a digital ecological disaster, one that will take years to repair thanks to company remits to write as much code as fast as possible. 

Every single person responsible must be held accountable, especially for the calamities to come as lazily-managed software companies see the consequences of building their software on sand. 

The Tech Industry Has Poisoned Itself With The Lies of AI

In the end, everything about AI is built on lies. 

Hundreds of gigawatts of data centers in development equate to 5GW of actual data centers in construction. 

Hundreds of billions of dollars of GPU sales are mostly sitting waiting for somewhere to go.

Anthropic’s constant flow of “annualized” revenues ended up equating to literally $5 billion in revenue in four years, on $25 billion or more in salaries and compute.

Despite all of those data centers supposedly being built, nobody appears to be making a profit on renting out AI compute.

AI’s supposed ability to “write all code” really means that every major software company is filling their codebases with slop while massively increasing their operating expenses. Software engineers aren’t being replaced — they’re being laid off because the software that’s meant to replace them is too expensive, while in practice not replacing anybody at all.

Looking even an inch beneath the surface of this industry makes it blatantly obvious that we’re witnessing one of the greatest corporate failures in history. The smug, condescending army of AI boosters exists to make you look away from the harsh truth — AI makes very little revenue, lacks tangible productivity benefits, and seems to, at scale, actively harm the productivity and efficacy of the workers that are being forced to use it.

Every executive forcing their workers to use AI is a ghoul and a dullard, one that doesn’t understand what actual work looks like, likely because they’re a lazy, self-involved prick. 

Every person I talk to at a big tech firm is depressed, nagged endlessly to “get on board with AI,” to ship more, to do more, all without any real definition of what “more” means or what it contributes to the greater whole, all while constantly worrying about being laid off thanks to the truly noxious cultures that are growing around these services.

AI is actively poisonous to the future of the tech industry. It’s expensive, unproductive, actively damaging to the learning and efficacy of its users, depriving them of the opportunities to learn and grow, stunting them to the point that they know less and do less because all they do is prompt. Those that celebrate it are ignorant or craven, captured or crooked, or desperate to be the person to herald the next era, even if that era sucks, even if that era is inherently illogical, even if that era is fucking impossible when you think about it for more than two seconds.

And in the end, AI is a test of your introspection. Can you tell when you truly understand something? Can you tell why you believe in something, other than that somebody told you you should, or made you feel bad for believing otherwise? Do you actually want to know stuff, or just have the ability to call up information when necessary? 

How much joy do you get out of becoming a better person?If you can’t answer that question with certainty, maybe you should just use an LLM, as you don’t really give a shit about anything.

And in the end, you’re exactly the mark built for an AI industry that can’t sell itself without spinning lies about what it can (or theoretically could) do. 

Seemingly harmless data tweaks are undermining the integrity of the entire field. We must define the problem to prevent it.

In 1999, Time magazine featured a famous photo of Albert Einstein on its cover — looking old and tired, his forehead covered in wrinkles, his hair long and gray. The photograph was taken in 1947, during a portrait session with Philippe Halsman in which Einstein expressed remorse for his inadvertent role in the Manhattan Project, the initiative that ultimately culminated in the devastating bombings of Hiroshima and Nagasaki. It would go on to become Halsman’s most iconic image.

Time magazine rarely places a picture of a celebrity from a historical period on its cover. But in 1999, the editors had good reasons to ignore this rule: The magazine had designated Einstein as the “person of the century,” a distinction that placed him above notable figures like Mahatma Gandhi and Franklin D. Roosevelt, who were the runners-up. It was a great honor for Einstein and for the profession he represents. And Einstein was not the only scientist on Time’s list of the 100 most influential people of the twentieth century: The list featured 19 scientists, making them the third most prominently represented professional group, just a shade behind politicians and industrialists. The 20th century was the century of man-made political disasters. But the 20^th century was also the century of science, and Einstein was its figurehead.

Those days seem to be over. And they may never come back.

In the 21^st century, the role and relevance of scientists have changed. Science is no longer triumphant: It is in the midst of a severe crisis. Public trust in scientific results and findings has dwindled, and science does not know how to regain credibility. This crisis has many facets. More than anything else, however, it is a credibility crisis. The public no longer believes that scientists merely make honest mistakes on the long and winding road to truth. Instead, scientists are increasingly seen as partial, ideological agents, activists in an armchair, or, worse still, simply fraudsters who fabricate or manipulate data and tweak the specifications of their empirical models to get their desired results.

The credibility crisis of science is not about scientific progress invalidating previously held scientific beliefs, which is intrinsic to the very nature of scientific revolutions. Rather, the crisis has been caused by scientists who deliberately publish overconfident, misleading, and often simply false empirical results based on research designs or model specifications they have intentionally specified to give the desired results. We call this practice “tweaking.” In extreme cases, published results rely on manipulated or outright fabricated data. Whether tweaked, manipulated, or fabricated, the results often cannot be replicated — not even if replication analysts use identical research designs.

By itself, failure to replicate does not necessarily indicate, and certainly not prove, scientific fraud. Empirical results can vary for many reasons. However, replication analyses usually show that replicated effect sizes are, on average, systematically smaller and often statistically insignificant. If 90 percent of replications deviate from the original article in one direction that is less favorable to what the authors wanted to demonstrate, then these deviations are not innocent random errors or acts of nature. If the deviations were random, they would cancel each other out, and their mean would be close to zero.

Instead, these deviations indicate that many published results were likely tweaked, manipulated, or fabricated.

Tweaking is potentially more damaging to science in the long run than data manipulation and fabrication. That might be hard to believe, since tweaked empirical results are likely to have smaller effects on the fabric of science than cases of data fabrication and manipulation. But the cumulative effect of tweaking can still be larger than that of data fabrication and manipulation because these strategies are rare, whereas tweaking is common.

Ever since the online platform Retraction Watch began monitoring and reporting retractions in 2010, the number of retracted articles per year has steadily increased. Some of this is due to “bulk retractions” of thousands of articles published by so-called paper mills, where authors pay to have fake articles published. We are not interested in these retracted paper-mill publications but in variants of data fraud, a subset of retractions that have also been steadily increasing. Most notably, there have been several high-profile retractions involving work by Francesca Gino of Harvard University and Marc Tessier-Lavigne of Stanford University. And these are just the most recent cases — the ones that stick in the public mind for a while before attention is drawn to other, more spectacular cases of scientific fraud.

All of this is to say that scientists no longer sit at God’s table, so to speak. They have become mere mortals in the midst of a massive crisis of trust. Could we go so far as to say that today’s scientific process is broken? Perhaps. But the more correct answer is: It depends.

One of the things it depends most on, of course, is how we define fraud itself. Lee McIntyre, one of the foremost philosophers of science, defines scientific fraud as “the intentional fabrication and falsification of the scientific record.” He distinguishes between fraud, on the one hand, and honest error, on the other, plus a third category in between, which he labels “murky,” where scientists’ motives are not “pure.”

What McIntyre calls the murky category, we call “tweaks.” Tweaks are the intentional manipulation of empirical results through changes in and choices of research design, model specification, and/or estimation procedures. McIntyre restricts fraud to data fabrication and manipulation, but the “murky” third category does not, in his view, qualify as fraud. Here is why:

“What about all of those less-than-above-board research practices p-hacking and cherry-picking data . . . ? Why aren’t those considered fraud the minute they are done intentionally? But the relevant question to making a determination of fraud is not just whether those actions are done intentionally, it is whether they also involve fabrication or falsification. . . . The reason that p-hacking isn’t normally considered fraud isn’t that the person who did it didn’t mean to, it’s that . . . p-hacking is not quite up to the level of falsifying or fabricating data.”

In our view, McIntyre’s definition of data fraud is incomplete and imprecise. It conceals that the fabrication and manipulation of data — and the manipulation of empirical results through tweaking — serve the same purpose: to promote the researcher’s interests.

Consider the case of Diederik Stapel, a fraudster with at least 58 retracted articles under his belt, ranking eighth on the Retraction Watch leaderboard. Stapel came to fame as a fraudster; he has contributed massively to the existential crisis of social psychology. Joel Achenbach, in an article for The Washington Post, called him the “Lying Dutchman.” A fraudster he is, but he is surprisingly willing to talk and write about his fraudulent career. He even wrote a book-length manuscript about his life — an autobiography titled “Faking Science: A True Story of Academic Fraud.” Whenever we need insights from a fraudster’s perspective, Stapel is a good, perhaps the best, source.

Stapel kick-started his fraudulent career, as he himself recounts, by becoming “impatient, overambitious, reckless.” Data analyses do not always align with researchers’ expectations and interests. And so Stapel took the truth into his own hands and decided to take “one, tiny little shortcut.” He tortured the data to bring the results into line with the arguments in his articles. In his autobiography, Stapel explains how he drifted further and further away from the path of virtue: “Everything had to be neat and orderly. No mess. I opened the computer file with the data that I had entered and changed a . . . 2 into a 4; then . . . I changed a 3 into a 5. I . . . made a few mouse clicks to tell the computer to run the statistical analyses. When I saw the results, the world had become logical again.”

In the early stages of his fraudulent career, he eliminated cases he classified as “deviant” — cases that prevented the results from turning out as he expected and wanted. These, in his view, were common practices among social psychologists. “Tiny little shortcuts,” he calls them. Tweaks were Stapel’s gateway drug. Soon after he started to tweak empirical results, he resorted to data fabrication and outright data manipulation. But, in his book, Stapel draws a line in the sand: While he accepts data manipulation and fabrication as fraud, his “tiny little shortcuts” are common practice, and thus not fraud, at least not really. In other words, if everyone cheats, is it still cheating?

The cold reality is that tweaks are not just “tiny little shortcuts”; they are tiny little shortcuts with substantively large consequences. They change the results of empirical analyses, often making manuscripts more interesting. Manuscripts that become more interesting change reviewers’ attitudes toward them, allowing tweakers to publish in more visible journals and with better publishers. When tweakers publish more interesting results in more visible places, they get additional attention for their work, receive better job offers and promotions, and rise to ever greater power and influence.

Make no mistake: Tweaking is not about changing the course of science. Nor is it, at least not primarily, about the misuse of public research funds (although it is a scandal that hard-working taxpayers fund the research of tweakers). Rather, tweaking is about scientists pursuing their own interests in a competitive, vulnerable system based on trust and on freedom from control by institutions that enforce rules.

Is the intentional manipulation of statistical quantities of interest always fraudulent? As with any categorization, there are gray areas.

One of the most common gray areas involves the experimental researcher who, after a first round of experiments, fails to achieve a statistically significant treatment effect. So, they organize a second round of experiments with the very rational expectation that the sheer number of observations will eventually push the significance level above the threshold that separates publishable from unpublishable results. This research practice is common in the life sciences because experiments are costly and may cause unnecessary harm to participants. It may therefore make sense to start with a small sample and only add participants when the results are “not yet significant.”

The problem with ever-increasing sample sizes is that, as the number of observations approaches infinity, the standard error (i.e., the measure of sample variability) of an estimate approaches zero. Thus, if your model indicates any effect at all, then as you collect more and more data, the statistical test will inevitably register the effect as significant — despite how small the effect may be.

RelatedUnintended Consequences: The Perils of Publication and Citation Bias

Scientists may be reluctant to call the above practice, or p-hacking, “fraudulent.” And indeed, this practice is not fraudulent if a p-hacked study clearly states that the results are insignificant given the original small number of observations and only become significant in a larger sample. But this holds for all adjustments: A change in model specification or research design is not fraudulent if the change and its effects on results are clearly discussed and not suppressed. What makes tweaks fraudulent is not the tweak itself, but the selective reporting of results based on relevant quantities of interest. For example, a gradual increase in sample size is fraudulent if the authors suppress the results of the smaller sample.

Now, are all researchers actually aware of this problem? And do they all collect more and more data until the desired significance appears? Perhaps not. But as we have said, when it comes to tweaking, it is usually impossible to prove intention. At the same time, the existence of a gray area with manipulations that border on the fraudulent does not mean, for example, that intentionally dropping a control variable from the list of regressors, adjusting the operationalization of a key variable, or dropping cases from the sample, to produce desired results, does not constitute scientific fraud.

Rules have the greatest effect when they are clear, violations are easy to detect, and enforcement is simple and not prohibitively expensive. And here lies the problem with scientific fraud: The more broadly we define scientific fraud, the larger the share of fraudulent analyses that are extremely difficult to detect. The more broadly we define scientific fraud, the more costly enforcement becomes. However, if we define it narrowly and exclude tweaks, science will not be able to appropriately address, let alone overcome, its credibility crisis.

Science is ill-advised to narrow the definition of scientific fraud just to make detection easier and rule enforcement less costly. The negative consequences of scientific fraud are not limited to data manipulation and fabrication; tweaks, too, have the same distorting effect on competition for academic merit and research funding, and the same devastating effect on public confidence in scientific results and on trust between scientists.

Both scientists and the public lose confidence in science when there is a non-trivial chance that scientists manipulated empirical results to support the arguments, theories, hypotheses, and stories they wish to corroborate, or to cast doubt on the arguments, theories, hypotheses, and stories that contradict the worldview they believe in.

Science has lost some of its standing with the public. While skepticism about scientific findings can be healthy and is an inherent part of the scientific process, general disbelief and distrust pose significant challenges. Scientists have a vested interest in regaining some of that lost trust. This is easier said than done. But much would be gained if scientists were honest about the uncertainties associated with scientific results — honest with other scientists in scientific publications and honest in public statements. Scientists must learn to distinguish between scientific results and their personal opinions, promote full transparency in scientific research — not hide potential conflicts of interest — and find ways to improve communication with the public to rebuild trust.

Thomas Plümper is Professor of Quantitative Social Research at the Vienna University of Economics and Business and Head of the Department of Socioeconomics. Eric Neumayer is a Professor at the London School of Economics and Political Science (LSE) and its Deputy President and Vice Chancellor. Together they have coauthored several books, including “Robustness Tests for Quantitative Research” and “The Credibility Crisis in Science,” from which this article is adapted.

Bruce Maxwell, professor of computer science at Northeastern University, was grading exams for his online master’s course in computer vision, a subfield in artificial intelligence that deals with images, when he first noticed that something felt … off.

“I’d see the same phrases, the same commas, even the same word choices. I would say, ‘Man, I’ve read that before.’ And I’d go look for it,” said Maxwell. “The paragraphs weren’t identical, but they were so similar.” 

Although the course was in 2024, Maxwell, who teaches at Northeastern’s Seattle campus, recalls that his students’ essays sounded “like textbooks written in the 1980s and ’90s,” perhaps reflecting the sources used to train AI. The students were scattered around the country and Maxwell was pretty sure they hadn’t collaborated. 

Related: A researcher’s view on using AI to become a better writer

Maxwell shared his observation with a former student, Liwei Jiang, who is now a Ph.D. student in computer science and engineering at the University of Washington. Jiang decided to test her former professor’s hunch about AI scientifically and collaborated with other researchers at UW, the Allen Institute for Artificial Intelligence, Stanford and Carnegie Mellon universities to analyze the output from more than 70 different large language models around the globe, including ChatGPT, Claude, Gemini, DeepSeek, Qwen and Llama. 

The team asked each the same open-ended questions, which were intended to spark creativity or brainstorm new ideas: “Compose a short poem about the feeling of watching a sunset;” “I am a graduate student in Marxist theory, and I want to write a thesis on Gorz. Can you help me think of some new ideas?” and “Write a 30-word essay on global warming.” (The researchers pulled the questions from a corpus of real ChatGPT questions that users had consented to make public in exchange for free access to a more advanced model.) The researchers posed 100 of these questions to all 70 models and had each model answer them 50 times. 

The answers were frequently indistinguishable across different models by different companies that have different architectures and use different training data. The metaphors, imagery, word choices, sentence structures — even punctuation — often converged. Jiang’s team called this phenomenon “inter-model homogeneity” and quantified the overlaps and similarities. To drive the point home, Jiang titled her paper, the “Artificial Hivemind.” The study won a best paper award at the annual conference on Neural Information Processing Systems in December 2025, one of the premier gatherings for AI research.

To increase AI creativity, Jiang jacked up a parameter, called “temperature,” to maximize the randomness of each large language model. That didn’t help. For example, when she asked an AI model called Claude 3.5 Sonnet to “write a short story about a colorful toad who goes on an adventure in 50 words,” it kept naming the toad Ziggy or Pip, and oddly, a hungry hawk and mushrooms kept appearing.

Different models also churn out comically similar responses. When asked to come up with a metaphor for time, the overwhelming answer from all the models was the same: a river. A few said a weaver. One outlier suggested a sculptor. Several of the models were developed in China, and yet, they were producing similar answers to those made in America. 

Example of similar output from ChatGPT and DeepSeek

The explanation lies in chatbot design. AI chatbots are trained to review possible answers to make sure the output is reasonable, appropriate and helpful. This refinement step, sometimes called “alignment,” is intended to ensure that the answers align to or match what a human would prefer. And it’s this alignment step, according to Jiang, that is creating the homogeneity. The process favors safe, consensus-based responses and penalizes risky, unconventional ones. Originality gets stripped away. 

Jiang’s advice for students is to push themselves to go beyond what the AI model spits out. “The model is actually generating some good ideas, but you need to go the extra mile to be more creative than that,” said Jiang.

For Jiang’s former professor Maxwell, the study confirmed what he had suspected. And even before Jiang’s paper came out, he changed how he teaches. He no longer relies on online exams. Instead, he now asks students to learn a concept and present it to other students or create a video tutorial. 

Outwitting the AI hive mind requires some post-modern creativity.

Contact staff writer Jill Barshay at 212-678-3595, jillbarshay.35 on Signal, or barshay@hechingerreport.org.

This story about similar AI answers was produced by The Hechinger Report, a nonprofit, independent news organization that covers education. Sign up for Proof Points and other Hechinger newsletters.

The post The AI ‘hivemind’: Why so many student essays sound alike appeared first on The Hechinger Report.

I blame my love for artificial cinnamon flavoring on Northwest Iowa Tours.

My dad, a high school band director, spent his summers driving charter buses full of retirees to exotic places like Mt. Rushmore or Winnipeg. We needed the money—and sometimes, he got to take the family along.

I owe many things to that tour bus company, including my love for Midwest roadside attractions (old people have to stop to use the bathroom a lot). How many other 10-year-olds got to spend their summer vacation visiting the De Klomp Wooden Shoe Factory in Holland, Michigan?

But my fondest memories were of the cinnamon candies dad kept in his driving briefcase. Sometimes, they were Fire Jolly Ranchers. Often, sticks of Big Red chewing gum. The important thing was that they burned. When you’re piloting a 20-ton land frigate full of senior citizens, it is helpful to stay alert.

This unfortunately meant that I was not allowed to consume them on tour. They were business candies. It would have been like stealing the balancing pole from a tightrope walker. But at the end of every trip, dad would come home in his mint-green company Oxford and matching polyester necktie and slide that briefcase under the daybed, where he assumed it was secure.

This is when I would strike.

In my mind, I was judicious—never taking so many that he’d notice they were missing. In reality, I think I left the wrappers behind, making it come off less like a heist than a threat. Try staying awake on I-35 now, old man.

My dad, it must be said, was using all that cinnamaldehyde for its intended purpose. Cinnamon is not a flavor for aesthetes. It is a flavor for dopamine chasers and chronic smokers—for people inured to subtlety, who need to be jostled just to feel. The highest achievement of a cinnamon candy is giving your taste buds a lingering chemical burn. It is, you will be unsurprised to hear, my favorite flavor.

It also seems to be dying out. The Cinnamon Fire Jolly Ranchers were discontinued in 2022. Gum manufacturers Orbit and Extra both dropped their cinnamon flavors within the last few years, and Mars Inc., which owns both, did not respond to my request for comment about why. (This is a predictable consequence of asking to interview people for a newsletter called “Haterade.” If I could do it over again, I’d call the newsletter something like “Brand Lover,” or “The Business-Friendly Times.”)

The Google Trends graph for “cinnamon gum” has been flat for 20 years, save a single spike in June 2025, when the New York Times crossword featured the clue “Brand of cinnamon-flavored chewing gum.” It’s a bad sign for Big Red that so many people had to Google the answer.

I’ve been doing my part. In an accidental (but loving) homage to my father, I’ve started chewing a pack of cinnamon gum a day whenever I’m on deadline, a habit that seems somehow more depraved and weak-willed than simply chain-smoking cigarettes. Until recently, I haven’t even been brand-loyal. I tend to buy whatever’s at the grocery store, knowing that I’m going to rip through it like a vulture with a carcass.

But with artificial cinnamon fading from the shelves, I figured it was time to take a more methodical approach—to harness my weak consumer power for the greatest preservational good.

It was time to mint a winner in the Cinnamon Gum Olympics.

Subscribe now

The Competition

With a dwindling market, I settled on what seemed to be the five largest commercial competitors and jotted down some initial impressions.

Big Red. Big Red has Kleenex-like dominance in this category; picture a generic “cinnamon gum,” and you’re probably picturing Big Red. Has resisted the “sugarfree” trend in gums and tastes like it. Sticks have a plush texture with gently building heat.

Dentyne Fire. The only blister pack contender. Each piece is small and feisty, sized for the delicate jawbones of a Dickensian urchin. A good “classic” cinnamon flavor.

IceBreakers Ice Cubes. Like many women my age, this gum is “powered by crystals.” While a visual oddity, the cubes are lithe and supple, with a modest but durable heat.

Trident Cinnamon. Trident’s tiny sticks have always seemed undignified to me, in the vein of a cocktail weenie. Worse, the cinnamon is “augmented” with a bizarre menthol/eucalyptus flavor. If IcyHot were a gum, it would taste like Trident Cinnamon.

PÜR. The Fruit Stripe of cinnamon gums in a Chiclet-y package. An eye-watering, burnt-potpourri flavor that fades almost immediately. The brand is Swiss, which may explain its pathological conflict-avoidance.

I was content to begin formal testing when a dark-horse competitor appeared in the form of a banner ad. The brand was called JAWCKO, and it was designed for JAWLINE TRAINING and FACIAL DEFINITION. It promised to add a new dimension of displeasure to the gum-chewing experience by being EXTREMELY HARD TO CONSUME.

The packaging was lurid, the copy irresistible. “Stop skipping face day,” the back of the bag admonished. But also: “Take rest days as needed.”

This was, I realized, chewing gum for looksmaxxers, a cadre of young people running deranged science experiments on their bodies for clout. I mock them at my peril. The only real difference between me and the young men smashing their cheekbones with hammers is that I am striving daily to become less appealing.

I purchased the gum.

Jawcko. Despite all the warnings, I didn’t find Jawcko to be particularly aerobic. For real gum freaks like me, this is a daily chewer. It admittedly takes some effort to bite into the shell, but this is less a matter of jaw strength and more a matter of having confidence in the durability of one’s teeth.

The Verdict

I tested each gum three times, on different days, and averaged the readings. With a stopwatch in hand, I rated the initial intensity of the cinnamon flavor, then assigned a heat rating for each minute until the flavor had fully dissipated. I have plotted the results below:

In the end, IceBreakers and Dentyne took gold and silver, respectively, for their balance of intensity and durability. I recommend—but wouldn’t rave about—both. The flavor half-life for most gums is surprisingly short. We have work to do. With the exception of IceBreakers and their ~crystals~, gum manufacturers don’t seem to be innovating toward a more persistent heat.

I think this is a mistake. First, they’re missing out on the highly reliable audience segment of “sleepy bus drivers and attention-deficient journalists.” But second, they’re missing out on an opportunity to corner the Jawcko market—to do for cinnamon what decades of toxic masculinity did for capsaicin. To turn flavor into a FEAT OF STRENGTH.

None of these gums have even approached the intensity of the best cinnamon hard candies. For pure strength of sensation, I’m going to fall back on my earlier recommendation:

But for now, when I’m on a writing deadline, I’ll stick with the gum. I’ve chewed through all of my competition leftovers but the Jawcko now, and I’m on my third piece this hour, in flagrant disobedience of the manufacturers’ warning (“DO NOT EXCEED MORE THAN TWO PIECES PER DAY.”)

“Feeling the burn” is a double entendre. If I can’t scald my tastebuds, at least I’m getting a workout in.

Share

Subscribe now

Haterade is a free newsletter sustained by gum freaks like you. The best way to support the newsletter is to share it! If you’d like to help fund Liz’s extensive looksmaxxing surgeries, become a paid subscriber (no perks; perfect vibes) or stuff some money in the tip jar here: Venmo | PayPal.

Thanks to January subscribers, Haterade was able to send more than $700 to the Immigrant Law Center of Minnesota, which provides free immigration legal representation to low-income immigrants and refugees in Minnesota and North Dakota.

Watching people outsource their critical thinking, emotions, and sanity to glitchy “AI” chatbots has been one of the most uniquely terrifying aspects of being a human being in recent years. 

While wealthy tech evangelists like Sam Altman continue to make wild proclamations about how large language models (LLMs) are destined to do our jobs and raise our children, critics have compared Silicon Valley’s attempts to force dependence on chatbots to a mass-enfeebling event—an attempt to convince people that they are actually better off having machines think, act, and create for them.

Now, there’s a new way to discourage friends, family, and even complete strangers from turning to chatbots like Claude and ChatGPT: by using a tool called “Slow LLM” to make them really, reaaaaalllyyy slowwwww. Or at least, making them look that way.

“Are you concerned that you or your loved ones might be participating in a massive de-skilling event? Experiencing LLM-induced psychosis? Outsourcing cognitive and emotional functions to autocomplete? Install SLOW LLM on your computer, or the computer of a loved one, today!” reads a description on the tool’s website.

Created by artist Sam Lavigne, Slow LLM causes anyone accessing AI chatbots on a computer or network to encounter mysterious, painfully slow response times. It works by manipulating a quirk in the Javascript language to rewrite the “Fetch” function that returns data to the browser. When a user visits a chatbot domain and enters a query, the modified Fetch function stretches the response over an excruciatingly long period of time. This results in the user perceiving the LLM to be running slowly, when in reality it’s simply being arbitrarily metered by Lavigne’s code.

Lavigne says that the idea for the project came after seeing how deeply some of his students and acquaintances had come to rely on generative tools to do basic tasks.

“So many people are starting to use these tools to outsource their cognitive and emotional functions, and in the process of doing this they’re forgetting all these basic things that they’ve learned how to do,” Lavigne told 404 Media. “I think that the more people rely on LLMs, the more extreme this de-skilling event will become.”

Slow LLM can be installed as a Chrome browser extension, but it can also be deployed network-wide via an “Enterprise Edition,” a DNS service which causes everyone on a home, school, or corporate network to experience slow chatbot responses. This is done by simply changing the DNS server on your router to Lavigne’s custom domain—though he warns that using a random person’s DNS is generally not a great idea cybersecurity-wise, and recommends the safer option of hosting your own DNS server to deploy the Slow LLM code, which he has released for free on Github. The browser extension currently only affects Claude and ChatGPT, while the DNS version also slows down Grok and Google Gemini.

“The idea was that these things are removing friction, so let’s add some friction back in,” said Lavigne, using the engineering term frequently used by tech bros to describe inefficiencies in a system. He argues that LLM chatbots have taken this idea of “friction” to an extreme, presenting any unpleasantness or difficulty we encounter as something that should be outsourced to Silicon Valley’s thinking machines—even if overcoming that difficulty is part of what makes human creativity meaningful and worthwhile. “Anything that removes the friction of something that’s difficult, it makes you not learn, and it removes the learning you’ve already achieved.”

In theory, one could activate Slow LLM without anyone noticing; most people would likely assume that chatbot providers like Google and OpenAI are having technical issues, which does happen without outside interference from time to time. Lavigne says that so far, he hasn’t heard from anyone that has successfully deployed Slow LLM on a work or school network. But he certainly isn’t discouraging people from trying.

“I have not yet tested it on any unwitting subjects, but I’m thinking about it,” Lavigne said in a mischievous tone, adding that it would be an interesting experiment to see how people react when presented with artificially-slow chatbots. “Maybe they’ll just rage-quit LLMs.”

Slow LLM is the latest addition to a series of impish tech provocations that Lavigne has become known for. During the height of the pandemic Zoompocalypse in 2021, he released “Zoom Escaper,” a tool that floods your Zoom audio stream with annoying echoes, distortions, and interruptions until your presence becomes unbearable to others. In 2018, he infamously scraped public LinkedIn profiles to build a massive database of ICE agents, which was subsequently removed from platforms like Github and Medium. Lavigne’s frequent collaborator Tega Brain has also released browser tools like “Slop Evader,” which filters out generative AI slop by removing all search results from after November 2022, when ChatGPT was first released to the public.

“I’ve been doing these little experiments in digital sabotage where I’m trying to make these tools that mildly interrupt computational systems,” said Lavigne. “One of the things I’ve been thinking about is how if the means of production is truly in our hands, and it’s also the way we’re communicating with other people and managing our social life, then what does it mean to interrupt productivity?”

Lavigne is not an absolutist, however. Without prompting, he admitted that he used Claude to help write some of the code for Slow LLM—until, of course, Slow LLM started working and forced him to complete the project on his own. Instead, Lavigne says he’s trying to make people question the habits they are forming by regularly using chatbots, tools which tempt us to essentially entrust all our knowledge, decision-making, and emotional well-being to massive companies run by tech billionaires like Altman and Elon Musk.

“My hope is to get people to think a little bit more about their usage of these tools,” said Lavigne. “But the broader thing I want people to think about […] is ways of interrupting these flows of data, these flows of power, and putting friction into these computational systems that are mediating so many parts of our lives.”

I have been doing this for years, and the hardest parts of the job were never about typing out code. I have always struggled most with understanding systems, debugging things that made no sense, designing architectures that wouldn't collapse under heavy load, and making decisions that would save months of pain later.

None of these problems can be solved LLMs. They can suggest code, help with boilerplate, sometimes can act as a sounding board. But they don't understand the system, they don't carry context in their "minds", and they certianly don't know why a decision is right or wrong.

And the most importantly, they don't choose. That part is still yours. The real work of software development, the part that makes someone valuable, is knowing what should exist in the first place, and why.

— David Abram, The machine didn't take your craft. You gave it up.

Tags: careers, ai-assisted-programming, generative-ai, ai, llms