Thomson Reuters spent $40 million fine-tuning an open-weights model on its own archive and turned it into a product it sells to law and accounting firms. OpenAI, by the same guest's count, has signed $700 billion of contracts.
The whole industry is being valued on the premise that a bigger general model is a better one. Eli the Computer Guy argues the opposite is now happening, and that the companies with proprietary data are the ones who will capture the value.
"I mean, look, to be blunt, OpenAI is dead. It's done."
He is a recurring guest on this show and makes videos about the AI industry, including a running series on the executives leaving OpenAI — which is where the evidence for that sentence comes from.
I listened to the full interview so you can skip it. 28 minutes of audio, 13 minutes of reading.
Here are the 10 takeaways that matter.
👤 Guest: Eli the Computer Guy, a technology commentator and recurring guest on The Tech Report who makes videos about the AI industry, including a running series on the executive departures from OpenAI
🎙️ Host: Isaac Pound, who presents The Tech Report
📰 Published: 31 August 2026 on YouTube (The Tech Report)
🔴 YouTube | 🔗 Episode page | ⏱️ 28 min | ✅ Time saved: 15 min
Key Takeaways
He says the two big labs have split, and only one of them is chasing scale
Anthropic shipped a protocol for talking to robots that already exist, rather than building a humanoid one
Altman's thesis is eight years old and the contracts signed against it are the trap
One Oracle contract alone is $300 billion, against a company he says is now at $700 billion of commitments
First-mover advantage has been wrong for three decades and people still believe it
His examples are Lycos and Friendster, both of which found the ways not to build the product
Thomson Reuters got a working legal and accounting product for $40M, which it can plausibly earn back
The differentiator is not the model, it is data nobody else can train on
Fidelity built a 350M-parameter model that only reads financial documents; IBM trained Granite on its own 80 years of business email
He expects data owners to stop selling to the labs, the way the studios stopped selling to Netflix
The Anthropic valuation is where he stops defending it
A $2T target and a claimed $30T addressable market, which he calls beyond stupid
1. The two labs are diverging
Pound opened on where the frontier labs go if open-weight models are catching up on capability and beating paid models on efficiency.
The guest's first claim was that OpenAI and Anthropic are no longer running the same play. "Well, I think one of the important things to understand right now is that OpenAI and Anthropic are actually diverging."
His view of the frontier model is that it is a maturing product, not a widening moat. "I've argued the whole concept of the value of the frontier model — I think that value has been decreasing over the past couple years, basically the same way the value of smartphones has."
The evidence he gave for Anthropic's different direction is a hardware protocol, which he named as MHS. "Anthropic actually came out with an AI protocol for simply communicating with robotic devices that already exist."
The contrast he drew is with the humanoid robot everyone is building. His examples of what already exists are a dishwasher and the large installed base of industrial robots
His conclusion is that scale has stopped paying for itself, because "the bigger frontier models don't actually solve problems better for people now, and they're simply more expensive all the way around"
2. Altman's 8-year-old thesis
Asked why OpenAI keeps building bigger models rather than more useful ones, he traced it to a decision made years ago.
"I think the problem is that Sam Altman painted himself into a corner." He was careful to say this is not stupidity: a new technology has a couple of approaches that work and a thousand that fail
"It's kind of like Thomas Edison said with a light bulb: he found a thousand ways not to make a light bulb, and one way to do it."
The thesis was set roughly eight years ago and is well documented: more data and more hardware resources equal more intelligence
"And that's one of the reasons why you really hear out of Sam Altman about AGI or the singularity — because of this idea that he thinks the most valuable thing out there is simply more intelligence, more intelligent systems."
The thesis brought investor money, and the money became commitments. "I mean, he's got one contract alone with Oracle that's $300 billion, and that's just one"
He said the company was at one point going to sign $1.4 trillion in contracts, as he understands it, and put the current figure at "I think they're now down to $700 billion in contracts."
"He has gotten hundreds of billions of dollars in investment, theoretically, with a $1 trillion valuation, with the idea that more AI equals more better."
The open question he framed is whether the position can be unwound at all. "Can OpenAI pivot kind of the way Anthropic has, or are they kind of just stuck on this path, and the only way for them is over the cliff at some point."
3. First movers do not win
Asked what AI actually looks like if Altman is wrong, he started by attacking the assumption underneath the whole race.
He said first-mover advantage has been disproved for thirty years and is still believed. The theory is that whoever takes a vertical first monopolizes it
His counter-examples are the ones nobody remembers. "The problem is, we've seen — the first search engine I used back in college was Lycos." And: "Friendster was the first real popular social media site."
The mechanism is that the pioneer pays for the education of everyone behind it. "And what you find out is that the first mover comes in, they figure out all the problems, they burn a tremendous amount of money, again, figuring out a thousand ways not to build a light bulb."
What matters is that the next generation does not repeat the mistakes, and so does not spend the money on them
Applied to this cycle: "OpenAI may spawn the future of artificial intelligence, but it won't be OpenAI."
His first illustration is Anthropic itself: "Dario came from OpenAI. Dario literally said he didn't trust Sam Altman."
He read Anthropic's product history as a deliberate sequence — a coding product first, then workflow products, then the model context protocol for communication, then the robotics protocol
4. Thinking Machines' Inkling
His second example of an OpenAI alumnus building the alternative is Thinking Machines, which he said is run by OpenAI's former chief technology officer.
The product is a model designed from the start to be modified. "So the idea with Inkling is, what if you literally design a model to be tweaked? That's the entire point of it."
Fine-tuning normally means taking a general model and tweaking it for a purpose. This inverts the order
The second piece is the tooling that does the tweaking. "Tinker connects to Inkling in order for you to create a system to fine-tune your own model."
The category he expects this to create is bespoke rather than universal. "I think that's what we're gonna start looking at, more of this customized intelligence solution versus a utility type intelligence solution that Sam Altman's going for."
The distinction he kept returning to is between the value of AI and the value of OpenAI, which he said are two different things
5. A $40M model beats $30B
Pound asked whether the shift to bespoke models is already visible. The guest's answer was one company.
"What's really interesting is they just came out with a new model that cost them $40 million." Thomson Reuters took an open-weights model and fine-tuned it on decades of its own journalism and proprietary information
The reason it can do this and OpenAI cannot is what the labs were never allowed to read. "They trained on all the world's information that they could get their grubby little hands on" — some organizations put walls up, and that material was never in the training set
The output is a product, not a model. The model sits behind CoCounsel, sold to law firms and accounting firms
He flagged that the "fiduciary-level" description carries a trademark symbol, and said that makes it a marketing claim rather than a legal one
His contrast is what the general labs trained on. He noted Grok's training on Twitter and OpenAI's on Reddit, and asked what a business choosing a back-end model makes of that
The whole argument, though, is the arithmetic. "OpenAI signed $30 billion in contracts. Regardless of how good OpenAI is, how do you make enough revenue to pay that back?"
Against $40 million, he said, a buyer can see how the money comes back
6. Data is finally the oil
Pound pushed on whether that is replicable, since the Thomson Reuters archive is what made it work.
His answer is that the decade of data hoarding may finally have a purpose. He noted that before AI was called the new oil, data was, and that nobody knew what to do with the pile. "We might actually now have a use for all this data."
Fidelity is his example of how small this can go. "It's like a 350 million parameter model, and basically it's only tasked with essentially processing financial documents."
"You don't need your LLM to be able to give you Shakespearean poetry — you just needed to process documents, right?"
IBM's Granite is his example of training for tone rather than capability. "So again, what IBM did is they trained the Granite model basically on IBM's business communications for about the past 80 years."
The problem it solves is that a model trained on social media is a risk in business correspondence
The strategic shift he predicts is the streaming wars replayed on data. "It's like, oh, this is just a little bit of extra money — and then they started to realize that Netflix was eating their lunch, and they pulled it all in house, and they realized that the content was valuable to sell itself."
He noted that selling data to the AI companies is currently one of Reddit's revenue lines
He also said the media companies that licensed to OpenAI, the Wall Street Journal among them, got peanuts for it, and that it surprised him at the time
"Why sell OpenAI a data set for $10 million, when, if I do enough work to clean it up and turn it into a product, I can create some kind of business that's, you know, giving us $50 million per year in revenue?"
7. Routing needs the data
Pound asked how this differs from what frontier models already do when they route a prompt to a smaller specialist model.
He named the mechanism and explained it plainly. That is the orchestration layer, or a harness: a first model whose only job is to send the question to the right one
"So if you ask a question about history, send it to a model that understands history; if you're asking a question about math, send it to a model that understands math."
His point is that routing is not the scarce thing. OpenAI can route, but it cannot route to a model it has no data to train
The same architecture inside a specialist is where he thinks it goes. Thomson Reuters could build one model on law and another on accounting and route between them inside CoCounsel
"So that's why I'm saying, unless Thomson Reuters actually licenses their information to OpenAI, OpenAI simply can't do what Thomson Reuters does, because they don't have the data to train on."
8. The executives are leaving
Asked what role OpenAI ends up playing, he answered in six words and then gave his reasoning.
"I mean, look, to be blunt, OpenAI is dead. It's done."
His evidence is tenure, and he listed three. A chief revenue officer who had spent a decade at Slack or Salesforce and "She was at OpenAI for 8 months and bounced."
A vice president for American sales with a similar background, gone after five months
A head of data centers who had been at Google for a decade and Meta for five years, gone after about thirteen months
The inference he draws is about the initial public offering. If a real listing is coming, he said, the executive payout would not merely be wealth but generational wealth — and they are leaving anyway
"Sam Altman bet everybody's farm on the incorrect decision. And I don't think there's any way back for him."
9. A $2T ask is beyond sane
He had been building a case for Anthropic through the whole interview, and then stopped.
"Unfortunately, their valuations are getting beyond stupid at this point."
The specifics he objected to. "And they want a $2 trillion valuation." He said the company wants to bring its listing forward into September or October
"And one of their things is, they apparently are telling investors that the total addressable market for them is $30 trillion."
He separated the two things he keeps being asked to judge at once. There is a technology stack that makes sense and is valuable, and then there is the pricing of it, and he said they are not the same thing
The conditional endorsement stands. He credited Anthropic with starting from a product people wanted — Claude Code — and building outward from it, and said "I think Anthropic has a better potential for the future, if they can get those numbers somewhere close to sanity."
10. Anthropic becomes a utility
Pound's last question was where the model developers end up if every company fine-tunes its own model.
For OpenAI, he expects a salvage operation. "Basically, how do you take the pieces of a dead company and turn it into something valuable? Maybe they're just a massive NeoCloud at the end of the day."
His precedent is Meta: "So yeah, Meta's latest thing is, oh, look, we have a lot of computers, and nobody cares about AI, so we'll just be our own NeoCloud now."
For Anthropic, he expects the software-utility position. "So I think basically Anthropic essentially turns into this generation's Office 365"
His analogy is email servers: plenty of organizations run their own, and plenty use Gmail, and both have a place
The customer he describes is one without the institutional knowledge or the appetite to spend on racks of servers, who puts a credit card in instead
The specialist and the utility coexist in his picture. Thomson Reuters does not care about robots; a company that does will use the Anthropic protocol
Bonus Insights
His sharpest aside is about what the general models learned from. On xAI's justification for training on Twitter to capture real human communication: "And nobody other than Elon Musk thinks Twitter is real human communication."
He read the CoCounsel positioning as a signal about the buyer rather than the technology. Thomson Reuters, he said, understands what a fiduciary is and built for lawyers and accountants, which is a different exercise from building the most capable model
The episode's framing quote is the opening of the show, taken from the middle of his own answer: that Altman has backed himself into a corner
The interview covers no company outside the AI industry, and he made no market call — every judgment here is about business models rather than share prices
His bottom line is that the AI industry has two separate things happening inside it: a technology stack that is getting cheaper and more useful, and a set of financing commitments made on a thesis about scale that he thinks is already wrong — and that the companies sitting on decades of proprietary data, not the labs, are the ones positioned to turn the first into a business.
Products, Companies & Tools Mentioned
OpenAI (The subject of the interview: a $1 trillion valuation and, on his count, $700 billion of signed contracts against a thesis he says was wrong)
Anthropic and Claude Code (His example of a lab building outward from a product people wanted, now shipping protocols rather than only bigger models)
Oracle (Counterparty on the single largest contract he named, at $300 billion)
Thinking Machines Lab (Founded by OpenAI's former chief technology officer; its Inkling model is designed to be fine-tuned and its Tinker interface is how you do it)
Thomson Reuters and CoCounsel (The $40 million fine-tune on its own archive, sold to law and accounting firms — his central example)
Fidelity (Built a 350-million-parameter model that does nothing but read financial documents)
IBM and Granite (Trained on roughly 80 years of IBM's own business communications, so it writes like a company rather than like a forum)
Grok and xAI (His example of training data he would not want behind a business product)
Reddit (Currently selling data to the AI companies, which he expects data owners to stop doing)
Netflix (The streaming precedent: studios sold content cheaply until they realized it was worth more in house)
Meta (Cited as already pivoting its compute into a neocloud business)
Microsoft 365 (What he thinks Anthropic becomes: the default general tool you subscribe to rather than build)
If this was worth your time, send it to someone closer to the industry than you are.
Get the latest market chatter as it happens: