What's up with datacenters? And the Three Horse Race
Welcome to Cautious Optimism, a newsletter on tech, business, and power.
Happy Tuesday. I’m a bit mad about the Windsurf saga. Not only did a potentially public-scale startup get its head ripped off by Google for a few billion dollars, it appeared that the remaining asset would be left high and dry, forced to compete with the company that just ate its founding talent. Enter Cognition AI, the folks behind the AI developer Devin, which bought up what was left of Windsurf, including its extant ARR and now smaller team.
It’s a great move by Cognition, which picks up an erstwhile $82 million worth of annual recurring revenue and becomes, overnight, public-company size. But wasn’t it a bit better for us, the market, when Windsurf was competing with Cognition and both were fighting against Google?
Seeing potentially leading, future public companies get stripped for intellectual parts so that major tech shops can try to buy (retain?) market dominance is not the exact sort of creative destruction that makes me feel warm, or fuzzy inside. It feels more like incumbents bending the market to themselves through sheer dent of the gravitational pull of their bank accounts. Boo.
That said, if anyone wants to offer $2.4 billion million hundred thousand for this blog, we can chat at your leisure. — Alex
📈 Trending Up: Japanese yields … whatever this is … Kiro … the cost of tariffs … oil prices? … problems not going away … data centers in space … Chinese GDP …
Vellum: Vellum is a startup that wants to help companies build AI apps that actually work, instead of merely shining in demos. A good project, and one to keep an eye on. I asked if Vellum’s public LLM leaderboard was a good way for the company to drive awareness. It said in an email that the leaderboard has “been crucial for increasing [its] visibility.” Chalk one more point up for the ‘public and useful data is a great startup GTM strategy’ perspective’ if I may. Vellum also said that it is growing around 3-4x per year, and has SaaS-like margins, not a lower percentage that we hear some AI shops deal with.
📉 Trending Down: Cybersecurity … cybersecurity (again) … Russian concern? … French deficits … safety in Moscow … trade relations with Europe … farming subsidies …
AI for [X]
Yesterday the government doled out nine-figure contracts to a number of American AI companies. Google, OpenAI, Anthropic, and xAI will receive up to $200 million for their help in applying AI to the national defense. Other than that we know little.
But! The government trough is deep, as Palantir has shown, so it should be little surprise that xAI has built a new Grok program for governments. Just what OpenAI did back in June.
We’re seeing more industry-specific specialization from the largest AI shops. Anthropic has products for education and finance, for example. OpenAI is also opening up a finance arm. You get the idea. I fully expect that we’ll see more of this sort of work from major AI shops so long as they can garner large contracts; the question before founders is simply what part of the market is most ripe for growth and is sufficiently sheltered from Sam or Dario or Elon showing up and squashing their efforts.
What’s up with data centers?
For a while there you couldn’t turn around with a major technology company or neocloud announcing a massive new data center project. They fell like rain from the sky, blanketing the Earth in liquid capital ready to cool and congeal into silicon brains. All very neat and awesome.
But while market attention has wandered a bit, the rush is still on. To wit, from very recent news:
CoreWeave plans to invest up to $6 billion in Pennsylvania to build out hundreds of megawatts of data center capacity.
The company told Bloomberg that sees “continued demand for larger and larger installations […] Three years ago, there were 10 megawatt facilities. Then they became 100 megawatt facilities. Now the conversation is kind of gravitating toward the gigawatt scale.”
Meta put more detail onto its own plans, with its CEO saying that “SemiAnalysis just reported that Meta is on track to be the first lab to bring a 1GW+ supercluster online,” and that his company is “actually building several multi-GW clusters” with the first, named ‘Prometheus’ coming online next year. Each of these huge data centers will be about as big as Manhattan.
I recall this quote from Gavin saying that in time that data centers would become larger than Manhattan. He’s being proved more right, more quickly, than I thought possible. The big tech companies are rich, and they smell trillions in market value out there to capture. So, the shovels are busy.
Today, Alphabet pledged a $25 billion buildout on new datacenter capacity, billions to improve two hydroelectric power plants, and more. (There are other power deals occurring, but I don’t want to deluge you with every single news event!)
What to make of all that? That the race to win the AI future is not slowing. At all. It may in fact still be accelerating, from a spend perspective. That despite a seeming pause in AI adoption.
In an odd way, I think we can put some of the impetus for the above on xAI. Its own massive data center showed through Grok 4 that you can still muscle your way to smarter models — don’t shoot in my direction, we’re speaking directionally. So, if more compute still equals more better AI models, then everyone is going to need more compute. Not to mention that if we want inference prices to keep falling, we’re going to need more digital brains.
For TWiST this week, I took a look at the cost of the above. The major hyperscalers have tinkered with their depreciation calculations to extend the lifespan of their chips, racks, cables, and power work. By pushing out the expected lifespan of server spend, Microsoft, Meta, and Alphabet gave themselves more time to wring value out of that capex. And given that we do not think that the major tech giants of the world are committing wanton accounting fraud, they were not fronting.
Then Amazon went back on the trend earlier this year, and reduced the expected life of its data center matériel from six years to five. If another hyperscaler does the same, maybe we should worry a little about near-term profitability. Until then, let slip the spend of war.
POTUS today backtracked on some Nvidia chip bans regarding the Chinese market. Great news for Nvidia, and Alibaba and Tencent, all of whom rallied on the news. Of course, H20s are not GB100s, but they are still pretty darn powerful.
Get it, Kimi!
Inflation watch
Government data indicate that consumer prices rose 0.3% in June, good for a 2.7% gain on a year-over-year basis — ahead of 2.4% inflation recorded in May. The headline CPI number was 0.1% hotter than expected, with the market anticipating a +2.6% number compared to the year-ago month.
Stripping volatile food and energy prices, core inflation rose a slimmer 0.2%, good for an annual gain of 2.9%.
So, are we seeing the impact of tariffs in the modest increase in inflation recorded in June? It depends on who you ask. Economist Jason Furman thinks you can see “signs of tariffs in these numbers and that is only likely to grow,” for one. CNBC writes that while “evidence in June was mixed on how much influence tariffs had over prices, there were signs that the duties are having an impact.” The Journal agrees.
On the other hand, POTUS wrote in response to the numbers: “Consumer Prices LOW. Bring down the Fed Rate, NOW!!!” Choose your fighter.
Inflation ticking higher, with many seeing rising inflationary pressure, is precisely not when central banks tend to cut rates. Then again, POTUS probably thinks Erdogan was onto something. Hell, even the NRO thinks that present tariff policy is nuts.
State capitalism, redux
Remember when POTUS earned himself admin access to how U.S. Steel operates after a deal with Nippon Steel? Don’t think that the move was an intellectual aberration. POTUS has also stated that the United States is a place where he has ultimate, tariff-derived pricing power. From his Time interview:
We are a department store, and we set the price. I meet with the companies, and then I set a fair price, what I consider to be a fair price, and they can pay it, or they don't have to pay it. They don't have to do business with the United States, but I set a tariff on countries.
Those two examples make it clear that so long as power accretes to POTUS, the current government doesn’t worry much about economic orthodoxy; at times libertarian in its bent (gut the CFPB!), and at times seemingly impressed with North Korean economic planning.
Speaking of which, POTUS took a massive stake in MP Materials, a rare earths company, through the Department of Defense a few days ago. Now, I could argue this one either way — but keep in mind that we’re trend-watching, not trying to score actions against our personal politics all the time. You can read the MP release here, but the U.S. government now owns around 15% of MP, which is going to scale up and build more rare earth magnets.
Then Apple struck a big deal with MP Materials to the tune of $500 million to, the companies said, “supply Apple with rare earth magnets manufactured in the United States from 100 percent recycled materials.”
You may applaud Apple for its business savvy. Locking up rare earth-derived materials supply is a big deal in a world in which the underlying materials (neodymium, samarium) are privy to global trade barriers. Or, you could argue that Apple very much wants to curry favor with the government, which is presently taking its App Store monopoly apart. You choose. What’s awful is that we have to; free markets are only free if they are, and are not if they are not.
The Three Horse Race
Wrapping with a chart because you deserve it, here’s a fun one from OpenRouter, a service that offers access to a bevy of AI models, tracking which its users opt to use to calculate a sort of index of AI company prominence. See if you can spy the interesting datapoint (the ranking showed under the chart is for the current week):
Yes, there really are only three countries in the global AI race. The United States, China, and France, via Mistral. Viva la France?