An enterprise buyer who watched a frontier lab cut its model prices by 30% then watched its answers come back 30% longer, on the same work. Rich Swier, who runs a company that buys tokens in bulk, says that is the business model rather than an accident, and that the people paying for it are the heavy API customers subsidizing the $20 and $200 consumer seats. Most of the coverage of the AI shopping spree treats it as strategy. He and his co-host treat it as a pre-IPO story with a pricing scandal underneath it.
"And who's getting screwed and here's the dirty little secret is who's getting really, really screwed are the heavy users."
Swier founded raia, which sells a vendor-neutral harness for deploying AI agents across OpenAI, Google, Microsoft and Claude, so he sees the token bill from the buy side rather than the sell side.
I listened to the full episode so you can skip it. 35 minutes of audio, 18 minutes of reading.
Here are the 13 takeaways that matter.
🎙️ Hosts: Lee Dixon, who works with the global client base at raia and presents The AI Guys, and Rich Swier, the founder of raia, an agent-deployment harness for enterprises, and a tech entrepreneur and author on AI
📰 Published: 31 August 2026
🔴 YouTube | ⏱️ 35 min | ✅ Time saved: 17 min
Key Takeaways
A price cut is not a price cut when the output length moves with it
Swier says a 30% cut in model cost shows up as answers 30% longer on the API bill
The buyers who cannot switch are the ones carrying it
The acquisitions are being paid for in what he calls monopoly money, and that is the point
Valuations will not get higher, and today's target is next year's competitor
Being a model is no longer a business, so every lab is buying its way into a platform
He expects an AI winter, and not because the technology stalls
Compute, energy, permits and government oversight all bite at once
The model has stopped improving; the software wrapped around it has not
What people credit to a new model is mostly the harness built on top of it
An 18-month-old open-source model is good enough for most of what enterprises actually do
Which makes the frontier labs' pricing power the weakest part of the IPO story
AI has borders now, and European and UK buyers are the ones drawing them
Going public is the worst thing that could happen to these companies' cost structures
The minute the burn is public, he says the token price doubles
Nobody uses one AI tool, and the differences people perceive are branding
He uses about seven, and says raw API access makes them hard to tell apart
Revenue that goes from nothing to tens of billions in three years was not won on demand alone
His comparison is boiling a frog one degree at a time
The defensive move for a business is owning the layer between the model and the user
1. Buy Now, With Monopoly Money
Lee Dixon opened on the consolidation running through the AI industry ahead of the expected listings, and asked why the largest companies are buying everything in sight.
Swier's first answer is that the currency is inflated and the window closes. "I mean, now's the time to strike because valuations are not going to get any much higher than what they are today."
"Everybody's pumping so much money into AI and everything is so inflated that when you have capital, especially when it's not real money, when it's monopoly money, now's the time to buy acquisitions, even if you're overpaying for those acquisitions"
The second reason is defensive, and it has a clock on it. Wait too long and an acquisition target becomes a competitor, or a rival buys it and locks you out of the category
2. A Model Alone Is Not Enough
The strategic shift he describes is from selling a model to owning a platform. "I think a lot of the big guys, are starting to realize that being a model by itself is not enough and that they have to start building out a broader platform."
Coding is the piece everyone bought first, on his account, because it is the critical component of the platform. He put the Cursor acquisition in that frame, grouping SpaceX with Grok and xAI as he did it
"So Claude expanded their platform, with Claude code, which has been very successful for them." He said that is what let Anthropic pass OpenAI, an outcome nobody would have predicted eight months ago
OpenAI has the product and not the developers, and he blames the distraction. "And then you have, even OpenAI, which has Codex, which is very strong, but they don't necessarily have as big of a presence in the developer community as Claude Code has because, they kind of got a little bit lazy."
He named device-building as the distraction and said Sam Altman has taken his eye off the ball. He guessed OpenAI might buy Replit or a company like it
The hardware companies are running the same play from the other end. He cited Nvidia buying Hugging Face as a bet on open source as a vector of attack
The governance argument underneath it: frontier models will be regulated by the United States and others, while open-source models are harder to control because whoever hosts them is a different party each time
3. Justifying the Valuation
The platform story exists because the token story cannot carry the price. A lab going to market can say it makes money on tokens, on tools and on a platform that competes across several software categories
The comparison he reached for is a design tool with a public price on it. "And so you can argue that you point over to Canva and say, okay, well, they have an 80 billion market cap."
"Well, Claude design in six months could eat away, or could legitimately have a shot at eating, 30, 40, 50 % of that market cap."
The point of the exercise is that the numbers have to be stacked up somehow, because "They're coming out with like a really ridiculous valuation." Dixon's assessment was shorter: "Way overvalued."
Whoever lists first sets the ceiling for everyone else. "And it's like the first domino to fall is gonna try to set the market, right?"
Either the first one performs and validates the price, or every issuer behind it argues it deserves more
4. OpenAI Burned Its Partners
The cash burn is the number the IPO has to survive. As Swier put it, citing reports he had read, "but I mean, even with OpenAI, like they're projected from some reports I was reading to lose this year between thirty three to sixty billion, right? As far as cash burn goes."
The model quality argument has stopped selling, because the field converged. He said the smartest model matters less to end users than it did, describing the field as a Venn diagram that is almost a circle
"It's really like brand loyalty, who do I want to hang my horse to"
"I don't just need a model anymore."
He thinks the market share plan was wrong from the start. "OpenAI thought they had a very clear road ahead where they were going to have an 80 % market domination"
"It doesn't matter how you, know, and that's being very conservative, assuming that open source is going to eat 40, 50 % of the market." He said neither Anthropic nor OpenAI planned for that
The sequence of fallings-out is what he thinks turned the industry against the company. "So the first thing they tried to do was the reorg, which pissed off Elon."
Microsoft came next, and is now trying to exit the relationship, while OpenAI still depends on it for compute. Dixon added that Microsoft owns a large stake and, on his own estimate, "Yeah, and Microsoft owns a pretty large chunk and I think they generate about 70 billion a year in revenue from him"
Swier described the result as an ego-driven approach that pushed Anthropic and the partners OpenAI had annoyed together, and emboldened the competitor
5. A Market Split 30/30/30
Dixon put the endpoint as an even split, and Swier gave it a number. "30, 30, 30 at this point."
Dixon added the rest: "I you can throw some of the other laggards in there like xAI or Grok."
Swier does not think Grok competes for that share, and does not think it needs to. He sees it as an internal asset for Elon Musk's companies rather than a market product
"Owning your own AI platform within his group of companies makes sense." Tesla, SpaceX and X give it both a home and a content supply
Google is the awkward case: good models, price-competitive, and too big to turn. "But they're also not really known, for, they're such a big company that it's very hard to turn that battleship at the pace that an Anthropic and an OpenAI can."
6. Public Markets End the Burn
Losing money is currently an asset, and that is a private-company privilege. "That's like a good thing. That's like a badge of honor. Like, look, we're burning cash. Look how much we're spending."
The moment they list, he said, they are under the same constraints Google is
The worked example both hosts reached for is Meta. "He started burning cash trying to think like a startup and the market said, no."
The stock fell, and the lesson Swier drew is that a public company answers to shareholders rather than to its founder's conviction
The capital that funded the private phase came from outside the public markets. "They go to the Saudis." He named Japan and Masayoshi Son as the other stops, and summed the era up as "They've been able to live like the rock star startup life, right?"
His advice to any of these chief executives is not to list at all. "And this is another reason why I would, if I was CEO of any one of these companies, I would never go public."
The complication he conceded immediately is that they are running out of cash, and the public market may be the only place left to get it
7. An AI Winter Is Coming
He expects a slowdown, and he was careful to say it is not a verdict on the technology. The causes he listed are compute, energy and the state
"So it's not like a negative AI thing. It's just that the fact that we've burned through a lot of our resources very quickly and nobody was prepared for it. And then we're also facing some serious government oversight, which is going to slow things down."
"We have to build some nuclear power plants in order to support all this demand."
Demand is running ahead of supply, and energy bills are rising with it
Being public makes the slowdown worse rather than better, because it adds a reporting cycle to an industry that has been moving without one
Google's patience reads differently in that light. Both hosts cast it as the tortoise, and Swier said Apple's chief executive has been playing the same waiting game: satisfy the market, show a plan, and wait for the private companies to weaken
The asymmetry he keeps returning to is that Google has to produce profit after profit while the private labs do not
8. AI Has Borders
Dixon's contribution here is first-hand and geographic. From the client work he sees, "there has been this very clear line in the sand that a lot of teams, especially in the UK and Europe, have started drawing to where they don't want to be as hyper reliant on Western models"
The reasons he gave are control and politics rather than capability, and the alternatives being looked at include what those teams could build themselves and what could come from China
Swier read the same shift as a loss of novelty premium. "Like this winter that you're talking about is I think it's because people have become more informed and the sheen on just the shiny new model or what we've released in the last quarter, like that just doesn't pay dividends to everyone anymore."
What buyers ask about instead: "They want to know like how am I driving costs down potentially on how I'm operating, or how am I gonna get more value out of maybe a model you released six months ago that has less of a token consumption but is performing just fine for my use cases."
His conclusion is that AI now has borders, and that OpenAI, Google and Anthropic are unlikely to dominate the East, where open-source and European models take a large share
9. The Plateau Is the Harness
Dixon's argument is that today's frontier advantage is a temporary one. "And that might be a disadvantage today, but quite honestly, it's a disadvantage in kind of in a made-up market pressure, right because in 18 in 18 months the best open source model will be better than what the model is today"
Swier thinks the urgency to be on the newest model is manufactured. "Okay, maybe it was good enough for a six months ago. What's the big deal?"
"So, that lie is starting to unfold and it would have never unfolded, if they didn't get so greedy, the frontier models, right? They got greedy on, price and they and they got greedy on how they wanted to consume data."
The claim that matters most for anyone valuing a model company: the models have plateaued and the improvements are coming from elsewhere. "The thing that's improving that the, the thing that's falsifying the improvements in the model today are not the model themselves. It's the harness that they built around the model."
"So essentially, we had this massive, every two weeks, a new model came out." That cadence has stopped
What people credit to a new model, he said, is the software wrapped around it and the experience it produces, and they are misreading which layer did the work
A plateau is exactly what open source needs. "And I think a lot of this is going to start when we hit that plateau on the S curve, that's going to give the open source models time to catch up. And then you're going to, it's going to be indistinguishable for most use cases."
10. Nobody Uses Just One Tool
The differences consumers notice are personality, and they are engineered above the model. "And you can tell that they all have kind of their own personality, tone, the way that they even talk or the way that they respond. Very distinct."
"And the reason why is because they are manipulating this at a level above and beyond the LLM."
On the consumer side he expects preference to persist, and multi-tool use with it. "Nobody has just one AI tool. I use ChatGPT quite a bit. I use Manus. I use a lot. I use like seven."
On the agentic side, where his own company operates, the gap closes. Going straight into the model rather than through a vendor's harness produces results he described as broadly similar across providers, on internal tests his team is running as it adds open-source support
His illustration is a television showroom. At home you have nothing to compare against, so "I would be like, this is the most amazing TV in the world."
"You don't know the difference. You don't know what you don't know."
The blind test he would run if he ran an open-source company is to put ten people in a room, tell them they have unreleased early access, and hand them an open-source model more than a year old
"It's all crap. It's not true." He compared the connoisseurship people claim over models to judging art or wine
The branding around a launch, he said, is doing most of the work — the executive on stage announcing that the last thing is dead and this thing is the future
11. Prices Fall, Bills Rise
This is the mechanism the episode exists for, and it is a margin argument rather than a conspiracy. "And I think the fluff also comes back to this margin thing that we've noticed, which is, these models, especially on the consumer even commercial side, they're not built for optimization of token consumption."
"They're built to, hey, we've lowered our cost per token by X, but you're using three X the time amount of tokens this time. So like we're always going to line our pockets if we need to."
The worked version: "Absolutely. Yeah, all Claude has to do is, okay, we're going to lower our model costs by 30%." The answers on the API then come back, in his phrase, "are 30 % longer, right?"
The analogy is the advertising market, and he was explicit that his version of it is a joke at Facebook's expense. "So Facebook, it's Mark Zuckerberg has a dial in his office, right? And he goes, we got to make 2 % increase in revenue. Okay. What he does is it just goes click, click, and that just makes all the ads, the bid and ask for all the ads that you have to buy on Facebook, 2 % higher in cost."
Nobody feels it, on that account, because every buyer assumes demand for the keyword went up
Who absorbs it is the part with an investor consequence. "And who's getting screwed and here's the dirty little secret is who's getting really, really screwed are the heavy users."
His own company is one: not "a $20 a head, $200 a head shop" but a bulk token buyer generating billions of tokens through the API
"We're, we're buying tokens in bulk. So we're the ones getting screwed"
The reason they cannot leave is that they have built infrastructure on top of the vendor, and the reason the price moves is that the same vendor is subsidizing consumer subscriptions and needs the margin back
12. $0 to $80B in 3 Years
He does not think the revenue curve is explicable by demand. "But the reality is you don't go from $10 billion or even $0 in revenue to $80 billion in revenue in two, three years, right? Just because we really like talking to Claude, it just doesn't, it doesn't work that way."
"They double their revenue simply by ratcheting up, the cost and, finding that ceiling" — the ceiling being the point at which customers cancel
"It's like boiling a frog." And: "They turn up one degree, one degree, one degree until the frog boils. If you did that in any other industry,"
His conclusion is that "you would be in front of the Congress", and that "it's price fixing. And that's what's happening."
The comparison he makes is to structured finance. "It's the same thing Wall Street does with money, right? They create just derivative products. They shift money around. They create these assets and bloat them up."
"And I think that's probably the biggest risk that we have right now is that there's nobody really auditing," what actually drives the price of a token
He credited an executive at Palantir with making the same point in public without naming it. "And he's not wrong. He's saying like these guys are taking us to the cleaners and nobody has oversight."
Going public does not fix it, on his reading — it accelerates it. "The minute they go public, their token cost is going to double."
A company that lists at a very high valuation has to hold a growth rate, and the way to hold it is price rather than discounts. He expects increases dressed up as one and a half or two percent adjustments in markets where customers are already dependent
13. Own the Layer You Control
The defensive position both hosts land on is about ownership of the layer between the model and the user, and Dixon put it as a rule for any business buying AI
Dixon's failure case is compounding dependency. "They can all just keep cranking and twisting the knobs until I have so much, basically tech debt in that I can't support it and I go away."
Swier's version is that the ability to switch models has moved from a nice-to-have to a requirement, which is also what his own company sells, and he said as much
"Don't put all your eggs in one basket. Don't depend on one vendor and keep as much of that control out of the hands of people who are not motivated to save you money or to optimize your business."
The precedent he draws on is the advertising and social platforms. "If you're not the product, if you're not paying for the product, you are the product and," — "It's just like monetizing clicks. Now they're monetizing prompts."
"And it's a dangerous game if you are trying to build a business on top of it, which I think most people are trying to use AI. But if you don't control that cost structure, or at least don't have the choice to control it, yeah, you're going have a rough year next year."
Bonus Insights
The hosts open by complaining that their own channel keeps getting flagged, most often, Dixon said, for political advertising in Armenia, a country the show has never discussed
They also keep any music they play under seven seconds to stay clear of licensing, and mentioned asking their producer about adding songs
Swier's description of using a newer model inside a coding tool is a complaint about verbosity, not capability. "Like when I use Fable inside of Claude code, I honestly, sometimes I tell it, I have no idea what you're, what you're saying right now."
His theory is that the model performs intelligence — reporting that it looked into something and analyzed it — and that the performance is what users are reading as quality
Dixon's version: the more you talk, the smarter people assume you are
Swier's bottom line is that the AI shopping spree is a story about pricing power rather than technology: the labs are buying platforms to justify listing prices their token revenue cannot support, and the customers with the least ability to switch are the ones funding the gap.
Products, Companies & Tools Mentioned
OpenAI (Losing tens of billions a year on his account, short of cash, and the company he says made the enemies that emboldened its competitor)
Anthropic and Claude Code (The platform extension he credits with putting Anthropic ahead of OpenAI in the developer market)
Google (Price-competitive frontier models and too large to move quickly, but the only one already living under public-market discipline)
Nvidia and Hugging Face (The hardware side buying into open source as, in his words, a vector of attack)
Cursor (Bought because coding turned out to be the critical component of an AI platform)
Canva (The $80B market cap used as the yardstick for what a design tool inside a frontier lab might take)
Microsoft (A large shareholder trying to exit the relationship while OpenAI still needs it for compute)
xAI and Grok (A niche product on his reading, valuable inside Musk's group of companies rather than in the open market)
Meta (The cautionary tale — a public company that tried to burn cash like a startup and was punished for it)
Replit (His guess at the kind of company OpenAI ends up buying to answer Claude Code)
Manus (One of roughly seven AI tools he uses himself)
Palantir (An executive there, he says, made the pricing-oversight argument in public without naming it)
raia (The hosts' own company, which sells a vendor-neutral harness for deploying agents and is adding open-source model support)
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: