Last week the frontier opened its books. This week somebody undercut it.
Nvidia agreed to spend about $7 billion on a company whose main asset is the machinery for building AI models, and it plans to use that machinery to give models away for free. Anthropic, meanwhile, is preparing to file publicly for what could be the largest IPO in history, on the strength of a $65 billion revenue run rate that assumes companies keep paying premium prices for premium models.
In between those two facts sits AT&T, which cut the cost of some of its AI coding work by 56% by quietly routing the easy jobs to free open models and eating a 2% drop in quality.
That 56 against 2 is the number I keep coming back to. For three years the question was which model is best. This week the question turned into which model is good enough, and that is a much worse question for anyone selling tokens at the top of the market.
Late last week Poolside told its shareholders that Nvidia would pay $6 billion to license its model-building software, an internal system the startup calls Model Factory, and separately invest $1 billion at a $12 billion pre-money valuation. More than 100 Poolside engineers move over with it. The Wall Street Journal reported the details over the weekend, and the destination for all of it is Nemotron, Nvidia's family of open-weight models that anyone can download and run without paying Nvidia a cent for the model itself.
Read that once more. The most valuable company on the planet just spent seven billion dollars assembling the ability to produce frontier-adjacent models, and its plan is to hand them out. The stated target is China, where DeepSeek, Moonshot's Kimi K3, and Alibaba's Qwen have made downloadable models the default choice for a lot of the world's developers. The unstated target is closer to home. A capable free model competes with OpenAI and Anthropic, two companies Nvidia has invested in, sells to, and in OpenAI's case just backstopped to the tune of $105 billion on an Ohio lease.
The logic isn't complicated once you stop thinking of Nvidia as a chip vendor. Every free model that runs well on Nvidia hardware sells more Nvidia hardware. It's the CUDA playbook, which is to give away the layer that makes the expensive layer necessary. And the structure of the deal is now a pattern rather than a one-off, since Nvidia did roughly the same thing with Groq's inference technology and its engineers, and is reportedly circling Perplexity after weighing a licensing deal there too. Buy the technology, hire the team, leave the company standing, avoid the antitrust review that a real acquisition would trigger.
Here's why this landed harder than a normal funding headline. The same week, The Information reported that AT&T has cut costs on some coding work by as much as 56%, with about a 2% drop in quality, by running employee requests through a router that sends the easy ones to open models. Its internal platform handles roughly 45 billion tokens a day. Around 40% of employee AI requests already go to open weights, including Nemotron, Meta's Llama, and Google's Gemma, and the company wants that between 60% and 70%. Mark Austin, the AT&T vice president running this, said open models are now "just as good or better" than the frontier models sold a year ago.
Note the careful version of that claim, because it matters. Nobody is saying open weights have caught today's frontier. The claim is that last year's frontier is now free, and that most enterprise work does not need this year's. Goldman Sachs published a note arguing the cheap tier may end up being good news for the big platforms, on the theory that affordable inference means far more of it, filling the capacity everyone is borrowing billions to build. That's the optimistic read, and it might be right. It's also exactly what you'd expect the sell side to say about a buildout its clients are financing.
My take: Nvidia is doing something the labs cannot answer with a better benchmark. It is deliberately collapsing the price of the layer above its own, because it wins either way and its customers don't. A frontier lab has to defend a per-token price. Nvidia has to defend a per-GPU price, and free models drive GPU demand up, not down. So the company with the deepest pockets in the industry now has a financial reason to make intelligence cheap, and it just paid seven billion dollars to get better at it. If you sell tokens for a living, that is a strange and uncomfortable thing to watch your biggest supplier do.
On Thursday, Bloomberg reported that Anthropic expects its IPO to match or beat SpaceX's record. SpaceX raised $75 billion in June, which grew to $86.2 billion once the overallotment was exercised, the largest first-time share sale ever recorded. Anthropic has filed confidentially, is working with Morgan Stanley, Goldman Sachs, and JPMorgan, and could file publicly as soon as the end of this month, which is to say within days. CFO Krishna Rao has been briefing investors and pointedly not naming a valuation, though the number floating around private markets is roughly $2 trillion.
The growth underneath it is real and genuinely hard to argue with. Bloomberg put the annualized revenue run rate above $65 billion at the end of July, up from $47 billion in May and about $9 billion at the end of last year. Second-quarter revenue came in above $11.5 billion, with positive adjusted operating income, which I wrote about last week. Backers expect $100 billion to $120 billion by year end. There is no company in recent memory that has scaled revenue like this.
Then there's the other data set from this week. Ramp, the corporate card company, tracks AI spending across more than 70,000 US businesses, and its July numbers show Anthropic at about 43.5% of those customers against OpenAI's 39.7%. A real lead. But Ramp's economist Ara Kharazian says OpenAI has been growing faster so far this quarter, and the model-level detail is the part that should give a prospective shareholder pause. Fable 5, the flagship, accounted for roughly 6% of the tokens businesses bought from Anthropic last month and 11.4% of the dollars. OpenAI's GPT-5.6 Sol accounted for about 25% of its tokens. Kharazian's read on Sol is that it's "increasingly the choice for developers", while Fable struggled on price and on the data-retention rules regulators attached to it.
Ramp's sample skews toward tech companies and misses big enterprises that expense through other tools, so don't treat it as the market. Treat it as a texture check. What it suggests is that business AI spending is not sticky, that buyers move with each release, and that the most expensive model in the lineup is not where the volume goes. Add one more detail from The Information's reporting, which is that Anthropic is separately advising customers on how to burn fewer tokens in Claude Code. Helpful. Also an odd look eight weeks before you ask public markets to underwrite token growth forever.
There's a governance wrinkle too. The Information reported that Anthropic is weighing super-voting shares that would hand Dario Amodei, who owns about 2% of the company, and his co-founders outsized control after listing. You can defend that as protecting a safety-first mission from quarterly pressure, and Anthropic will. You can also notice that when Elizabeth Warren objected to Musk's 82% voting control at the SpaceX listing in June, the argument against it was the same one, and the argument for it was also the same one.
My take: I write this newsletter with Claude, so take the skepticism as calibration rather than a swipe. Anthropic's growth is the most impressive revenue curve I've ever seen, and I don't doubt the enterprise demand is real. What I doubt is the durability of the price. A record-setting listing asks buyers to believe that companies will keep paying $10 per million input tokens for the top model, in a week when AT&T showed it can route most of that work to free weights for a 2% quality hit and Nvidia announced it's funding better free weights. Anthropic's answer has to be that the hardest work still needs the best model and the hardest work is where the money is. That's a reasonable answer. It's also the entire bet, and the S-1 is where we finally get to check it.
Every argument in AI for the last year has been about the top of the market. Who has the best model, who can switch it off, who pays for the power, whose books look better. This week the interesting action was underneath all of that. A chip company funding free weights. A telecom quietly cutting its bill in half. A model hub suddenly worth $13 billion because it's where the free stuff lives.
The thing about a commodity floor is that it doesn't arrive as an announcement. It arrives as a procurement decision nobody blogs about. AT&T didn't declare independence from Anthropic. It installed a router, watched the invoices, and moved 40% of its traffic. Multiply that by a few thousand companies and the revenue curve everyone is about to buy shares in starts depending less on how good the best model is and more on how much better it has to be to justify five times the price.
I don't think this is the frontier losing. Somebody still has to build the model that everyone else's cheap version copies twelve months later, and that work is getting more expensive, not less. But the shape of the business is changing under the labs while they're on the road show. Two of them are about to list into a market where their most expensive product is the one customers use least, and where their own supplier is bankrolling the alternative. The frontier is still the frontier. It just stopped being the only place to get work done.
For months I've written about other towns doing this. Monday night it was ours. Just before midnight, after a two-hour hearing with 63 residents signed up to speak, the Durham County commissioners voted 4 to 1 for a nine-month moratorium on new and expanded data centers, running until May 2027. WRAL reported the timeline was set to line up with the city of Durham's own pause so the two can write rules together.
The details are worth knowing. The pause blocks new hyperscale facilities but lets enterprise and small projects under 100,000 square feet proceed, and commissioners amended it before the vote to add cooling and backup-generator requirements and to make applicants state their expected electrical load in writing. The lone no vote came from Commissioner Nida Allam, who thought it didn't go far enough. County staff say a draft ordinance could be ready by December. And as with every one of these, state law means it cannot touch projects already in the pipeline.
Zoom out and the pattern from the last few issues holds. Raleigh gave up on a statewide framework when SB 730 died, so the rules are being written one county at a time, on a meeting calendar rather than a docket calendar. Meanwhile the mechanism that actually decides who pays is still not moving. The Utilities Commission decision on Duke's Carolinas settlement is expected in November, and the large-load tariff proceeding, the one that would put data centers in their own rate class, still hasn't started. Durham County can decide where a building goes. It cannot decide whose power bill covers the substation.
Now the part that ties back to the top of this issue. If you build with AI here, the AT&T number is the most useful thing to come out of this week. A 56% cost cut for a 2% quality drop is not a research result, it's an engineering discipline, and it's exactly the kind of work the Triangle is good at. Put a router in front of your model calls. Measure quality per task instead of picking one vendor for everything. Keep a cheap open model warm for the easy 40% of your traffic. That habit also happens to be the best insurance against the last two months of headlines, because a stack that can swap models on a config change survives a price hike, an outage, or a government letter.
That's the week the cheap seats came for the frontier's price. See you next Wednesday.
Daniel
BullCity AI · Durham, NC
P.S. If you've put a router in front of your model calls, hit reply and tell me what share of your traffic goes to open weights and what it actually saved you. I'm collecting real numbers, because AT&T's are the only public ones I trust and one data point is not a trend.
P.P.S. Forward this to whoever signs your AI invoices. There's a good chance 40% of that bill is buying a frontier model to summarize a pull request.