The compute arms race goes public

00:00:04 SpaceX filed its IPO today, and buried in the S-1 is a number that doesn't make sense until you sit with it. Anthropic is paying $1.25 billion a month through May 2029 for compute capacity at SpaceX's Colossus data centers in Memphis. That works out to $15 billion a year — nearly double the $18.7 billion in revenue SpaceX reported for all of 2025.

00:00:30 One company's annual revenue has been contractually committed to another company's monthly capacity fee. The deal includes a 90-day exit clause for either side, which is sensible given how fast this industry moves, and probably necessary since Anthropic's Claude competes directly with X's Grok.

00:00:52 But the exit clause is the kind of detail that gets lost in headlines. Under the hood, this is a multi-year bet on a partner you could walk away from in three months. What's striking from the filing is the scale of SpaceX's AI spending. The company spent $12.7 billion on AI capital expenditures in 2025, about 61 percent of total spend.

00:01:18 In the first quarter of 2026, that figure was $7.7 billion on AI versus just $1 billion on the space division. The AI division lost $6.3 billion on $3.2 billion in revenue in 2025, and $2.5 billion on $818 million in the first quarter of 2026. SpaceX is spending more on AI infrastructure than space operations, and losing billions doing it.

00:01:44 Meanwhile, Anthropic is reportedly cruising toward its first quarterly operating profit, with revenue expected to hit at least $10.9 billion this quarter, more than double the $4.8 billion from the March quarter. On X, Musk wrote that SpaceX stands ready to offer similar deals to other AI companies.

00:02:07 He called it AI compute as a service at significant scale. A rocket company is positioning itself as a compute provider. That isn't a side bet anymore. It's a line of business with real P&L risk. Capacity is so constrained at the top that a deal this size, this unusual, or this asymmetric is actually possible.

00:02:30 The exit clause is the only thing keeping it from looking like a captive arrangement. For anyone building models, the signal is straightforward: compute access at this scale now requires partnerships that look more like infrastructure joint ventures than vendor contracts.

00:02:51 Nvidia's own quarter came in on the same day, and it tells the other half of the story. Revenue came in up 85 percent to $81.62 billion, alongside an $80 billion share buyback program and a raised dividend. On the face of it, this is a company at the peak of its run.

00:03:09 But in the same interview cycle, CEO Jensen Huang told CNBC he'd largely conceded China's advanced AI chip market to Huawei. China once accounted for at least 20 percent of Nvidia's data center revenue. The Trump administration told Nvidia in April that it would need a license to export chips to China and a handful of other countries.

00:03:33 Huang told CNBC's Sara Eisen that the demand in China is quite large, that Huawei is very strong, and that their local ecosystem of chip companies is doing quite well because Nvidia evacuated that market. He added that he doesn't have any expectation regarding approvals to sell advanced chips into the country.

00:03:55 That's a remarkable thing for the CEO of the world's dominant AI chip company to say publicly. Conceding a market doesn't usually happen with fanfare. It's usually a slow retreat behind compliance, pricing, and export control. In this case, the timeline is compressed.

00:04:14 The Trump admin's April directive forced the exit. The market that is filling the gap is one Nvidia helped build — Huawei's AI chip line. Reuters reported last week that a few Chinese companies, including Alibaba, Tencent, ByteDance, and JD.com, had received approval to purchase H200 chips.

00:04:34 But a U.S. trade representative said chip export controls weren't part of discussions during last week's China talks. The approvals are intermittent, not systematic. Huang also told CNBC that Nvidia would be more than delighted to return to the Chinese market if conditions improve.

00:04:54 He described the company's first priority as supporting its supply chain amid surging demand. And he said the idea of Nvidia becoming many times larger than it is now isn't out of the question. But the China loss is structural. It isn't a temporary dip. It's a market that is being rebuilt without Nvidia's architecture, and once that ecosystem solidifies, re-entry becomes harder regardless of what Washington decides.

00:05:24 The Google I/O keynote yesterday gave us a different angle on the same infrastructure question: what happens when the compute layer becomes the product? Google split its agent strategy into two lanes. On one side is Antigravity 2.0: a standalone desktop application, a CLI, and an SDK that lets individual developers build, define, and run hosted agents starting from a Gemini API key.

00:05:50 There's no Cloud console, identity roles, or billing setup. Just an API key and a prompt. On the other side sits the Gemini Enterprise Agent Platform: governed deployment with identity controls, audit logs, and policy enforcement for teams that need it. The agent definitions carry across both surfaces.

00:06:11 Same harness. Janakiram MSV at Forbes put the strategy clearly: what connects them is a single agent runtime layer that handles reasoning, tool calls, and code execution. The continuity is the differentiator. Amazon and Microsoft both offer strong agent tooling, but neither presents the same kind of consumer-to-enterprise API on-ramp that is visibly continuous with its enterprise platform.

00:06:38 The detail that matters for developers is the on-ramp. You can start in Google AI Studio with a Gemini API key. The key is associated with a Google Cloud project that Google creates in the background, but the builder never touches it. Managed Agents is reachable from that same key.

00:06:57 The friction isn't eliminated, as the reporting puts it, but deferred. Google is also encouraging Gemini CLI users to migrate to the Antigravity CLI. The Gemini CLI service for free, Pro, and Ultra users stops late June. Enterprise access continues under paid keys.

00:07:15 Developers who built workflows on the older tool now face a forced move, and consolidation tends to surface complaints about usage limits and changed defaults before it settles. The deeper open question is commercial. A low-friction entry point lowers the barrier to starting, but Google still needs those developers to convert into Google Cloud customers.

00:07:39 The funnel works if the developer experience stays frictionless long enough. The risk is that once the developer is hooked, Google's default path is toward Cloud billing, and that is where the trust test begins. There is also the preview problem. Managed Agents is in preview, and the documentation says features and schemas are subject to change.

00:08:03 Enterprise support is in private preview. A preview runtime is suited to experimentation, not production commitments. The platform continuity is clear; the immaturity is just as visible.

00:08:16 While Google is building developer funnels, Anthropic is building enterprise muscle. Bloomberg's Preeti Singh reported that the joint venture between Anthropic, Blackstone, and Hellman and Friedman, an unnamed enterprise services firm, has bought Fractional AI as its first deal.

00:08:35 Sources say Fractional is ending its OpenAI deal as part of the transition. This is a move from model API into the enterprise services layer, alongside private equity. It isn't just about selling inference tokens. It's about implementation, integration, and the long-running relationship that comes with it.

00:08:55 The venture structure itself is notable. Blackstone and Hellman and Friedman are both heavy-weight private equity firms with deep enterprise relationships. Anthropic brings the model capability and the customer base. Fractional AI brings the implementation practice and the client relationships.

00:09:15 The biggest labs are starting to see that enterprise AI adoption isn't a product problem. It's an implementation problem. And implementation requires people, process, and time — the kind of work that private equity firms know how to scale. This also reframes the Anthropic and SpaceX deal from yesterday.

00:09:36 Anthropic isn't just building a better model. It's building a company that spans compute capacity, model API, and enterprise services. That's a different kind of moat than the one OpenAI is building, which is still largely product-focused. Whether either approach wins depends on whether enterprise AI is a platform play or an implementation play.

00:09:59 The market is betting on both, and that means the infrastructure layer is becoming a competitive arena. The Information reported separately today that Anthropic is in talks to rent servers powered by Microsoft-designed AI chips. This adds a second infrastructure layer to the picture: not just SpaceX's Colossus, but Microsoft's custom silicon too.

00:10:23 Anthropic's Azure usage has been increasing steadily since November 2025. The company is diversifying its compute across three providers — SpaceX, Microsoft, and the remaining AWS and GCP mix — and that diversification itself is a signal of capacity anxiety at the top.

00:10:41 On the hardware side, there is a startup called Bolt Graphics making a GPU called Zeus that is designed for high-precision workloads. It features FP64-native vector cores and path tracing instead of the low-precision tensor optimization that has dominated the last few generations of Nvidia GPUs.

00:11:02 CEO Darwesh Singh told IEEE Spectrum that Nvidia has a fundamental lack of understanding of their customer. Jill Mueller, Bolt's CMO, put it more bluntly: Nvidia just throws stuff at you and there you go. Jacob Feldgoise at Georgetown's CSET confirmed the trend: a greater share of Nvidia hardware has been allocated to low-precision compute over time.

00:11:26 AI is sucking the computational units used for high-precision workloads out of that hardware. If you look at the highest-performance GPUs, the trend is clear. Zeus is fabricated on TSMC's older N5 process node. The company is betting that an older process node will keep Zeus competitive on price.

00:11:47 A rack with 28 Zeus GPUs claims real-time path-traced performance equivalent to 280 RTX 5090s at 4K resolution and 30 frames per second — a degree of accuracy required for professional rendering workloads that even the most graphically attractive games don't target.

00:12:06 The production risk is steep. Leading-edge silicon capacity is in short supply, and Nvidia has most of the leading capacity tied up. Bolt is taking the back seat on process node priority. But the gap it is targeting, geographic information systems, scientific computing, and industrial simulation, is under-served.

00:12:28 Nvidia's AI-first GPU design has created a genuine blind spot. High-precision workloads don't get the same silicon allocation anymore, and that creates an opening for a focused competitor. Whether Bolt can win on production volume, developer tooling, and customer mind share remains to be seen.

00:12:48 The architectural divergence is clear. Zeus supports rasterization, but only at about half the performance of a comparable Nvidia card. Bolt is explicitly choosing path tracing over rasterization. That's a bet that professional workloads will migrate to path tracing faster than the broader market expects.

00:13:09 If they are wrong, the GPU becomes a narrow tool in a wide market. If they are right, it is a category definer. The hardware story here ties back to the compute story. Nvidia is dominating AI inference and training, but the workloads that need precision are migrating elsewhere.

00:13:29 The infrastructure layer isn't one monolith. It is a set of overlapping bets on different kinds of compute, and Bolt is trying to carve a lane between them.

00:13:40 On a different note, Sean Hollister at The Verge spent yesterday building Android apps using Google AI Studio. He built three apps in one afternoon. One required just 148 words typed into a browser: a Doom-like text game, a calorie counter, and a workout tracker.

00:13:57 The demo works surprisingly well. You type words, hit install, and a working program appears on your phone via USB debugging. Gemini auto-generates code and deploys. You can iterate by describing what needs to change, and Gemini pushes a new version. Hollister's first reaction when confronted with Gemini's daily limit and upsell was: what if I try paying for a couple months?

00:14:23 That is unusual. Most demo users either dismiss the limit or accept the paywall. His instinct was to extend the experiment. But the apps themselves are rough. The text adventure had 11 rooms, a spam-button combat system, and secrets that glow with explanatory text instead of hiding.

00:14:42 The calorie counter asked the paid Gemini API for nutritional data instead of using local logic, which broke when Hollister didn't have a paid key. When he called out the error — 190 calories for a 16-ounce boba milk tea — Gemini caught it and fixed it. The daily limit hits fast.

00:15:00 The apps work for simple use cases but break on anything requiring real data or complex logic. The UX exposes every shortcut Gemini takes. It's a glimpse at where the tooling is heading, rather than where it is today. Stevie Bonifield at The Verge also made a personal workout tracker using the same tool and found it good enough to actually use.

00:15:24 The spectrum ranges from throwaway demo to functional tool, with a lot of apps landing somewhere in between. Vibe coding works as a prototyping layer, but it isn't a replacement for engineering. It's a first pass: fast, imperfect, and useful for exploration. The daily limit and the paywall are the constraints that will define whether it becomes a habit or a novelty.

00:15:48 There is also a subtle tension: the tools are getting better at generating code, but the constraint is becoming access — daily limits, API keys, paid tiers. The bottleneck is shifting from writing code to paying for the compute that writes it. That's a different kind of constraint, and it changes how the tooling evolves.

00:16:10 The infrastructure layer is shifting across multiple tracks. Seln Oriax.