AI capex is still accelerating. The contest underneath just got visible.

The five largest cloud builders now plan roughly $760 billion in combined 2026 capital expenditure — on company guidance compiled by the Financial Times and CreditSights, close to double 2025's $410 billion. The detail that matters is how Microsoft framed its own $190 billion: CFO Amy Hood attributed about $25 billion of it to component-cost inflation rather than expanded scope. Meta, lifting its range to $125–145 billion, cited component pricing too.

That justification language is what's interesting. If hyperscalers wanted to communicate strong demand, they would frame increases as "we are building more because customers need it." Instead, the framing is "we are paying more for the same plan." That is a subtle but meaningful shift. It does not say the demand isn't there. It says the demand isn't carrying the justification any more — the cost story is.

Both moves strengthen the capex-growth thesis on magnitude. The same evidence weakens the digestion bear thesis. And the framing shift itself strengthens the infrastructure-strain thesis — visible strain has now entered the language hyperscalers use, not just the margins they earn.

The Q3 earnings cycle in late October will tell us which way it goes from here.

SpaceX's S-1 filing on 20 May disclosed two things together. The first is the IPO itself — a reported target near $1.75 trillion (some accounts put it closer to $1.5 trillion), expected mid-2026 and on track to be the largest listing in market history. The second, less headlined, is a compute agreement with Anthropic worth about $1.25 billion per month, running through May 2029, for the full 300 megawatts of xAI's Colossus 1 data centre in Memphis. Total contract value runs upward of $40 billion.

That is the largest single AI compute deal ever disclosed, paid by the lab that has been most publicly disciplined about its unit economics. Anthropic's $4.8 billion quarterly revenue is one piece of evidence about where the cost-curve sits at the top of the market. Anthropic choosing to pay xAI $15 billion a year rather than build comparable capacity itself is a much sharper piece of evidence. The most cost-disciplined buyer in the market has revealed that building independently is not the cheaper path.

This connects directly to the rest of the physical-layer picture. xAI's on-site gas-turbine generation at Memphis — fast to deploy, and sidestepping the multi-year grid-interconnect queue (the wait to connect new generation or load to the grid) — is what makes Colossus's cost structure work. Separately, NextEra Energy and Dominion Energy announced a $67 billion all-stock merger, explicitly framed around AI data centre power demand. Same constraint, different repricing.

Worth holding one quiet thing alongside the strategic read. The xAI Memphis facility is the subject of an active NAACP lawsuit over alleged Clean Air Act violations from its on-site gas turbines. And the wholesale price of grid capacity across the PJM region has roughly doubled across two successive auctions, driven in large part by data-centre demand. Only a small share of that reaches households — an estimated 1.5 to 5 percent on retail bills, unevenly spread — but the size of the number is not really the point. The cost-per-watt advantage some operators have built is being borne, in part, by ratepayers who did not sign up to subsidise it, and that becomes its own risk vector in the PR tug-of-war: a natural consequence of a shift this consequential. The discourse has not yet caught up. At some point it will.

Two postures sit side by side this week. The East: DeepSeek's V4 Pro posts near-frontier benchmark scores at a fraction of the cost of US flagships — on published pricing its output tokens list several times cheaper. The West: Anthropic disclosed $4.8 billion in Q1 revenue, projected $10.9 billion for Q2, and — on the same fundraise projections — a first operating profit of about $559 million in that quarter, while cautioning it may not stay profitable across the full year as datacentre spending ramps. Two cost structures, two strategies, two demonstrations that frontier AI economics can work — through very different mechanisms.

Anthropic's profitability projection is the analytically richer data point. It proves that customer concentration on capability superiority can deliver Western unit economics in this specific window — when enterprise testing and adoption of intelligence layers happened to coincide with Anthropic's lead. The validation is real and the timing was good. The question is whether the concentration is loyal — or whether the market grows fast enough that loyalty matters less than absolute capture.

The case for loyalty mattering: Anthropic's pricing power depends on customers staying. Cost pressure from the East and capability pressure from other Western labs both threaten that.

The case against: forecasters expect AI token consumption to rise by an order of magnitude or more by 2030, and in agentic-coding workflows token spend is already becoming a material line against engineering payroll. If the market grows that fast, every credible Western lab gains customers in absolute terms even as relative share fragments — a rising tide carrying all boats, but riskily, since the rise itself depends on enterprise willingness to keep spending into uncertain ROI.

This loops back to the cost-per-watt thesis. If the market grows as forecast, the watt-per-watt question matters most on the supply side — who builds compute cheapest captures the largest absolute slice of an expanding pool. If growth disappoints or fragments, the question moves to the demand side — who keeps customers. Anthropic's profitability proves the model works in this window. Whether it works at the scale forecasters project is the unresolved question the publication should track from here.

Tesla's unsupervised robotaxi service now spans three Texas metros — Austin, plus Dallas and Houston, which went live on 18 April — and Musk has reiterated a year-end US-wide ambition. The footprint is the headline; the fleet behind it is the reality check. As of late May, crowdsourced trackers and Texas DMV records put the total unsupervised fleet across all three metros at roughly twenty vehicles, inside geofences measured in tens of square kilometres, with a reported crash rate well above the human baseline. (Full Self-Driving software subscriptions — Tesla's paid driver-assist tier — number far higher, but those are not deployed robotaxis and do not count toward this thesis.)

So the honest read is narrower than the bullish one. The category is proving out — Waymo is running around 500,000 paid rides a week across ten-plus cities — but the Tesla-specific bet this thesis tracks is still a small, geofenced pilot, not deployment at scale. Three named metros counts toward the five-metro threshold only in the most literal sense; paid-ride volume, not metro count or FSD subscriptions, is the working-definition test. On this week's evidence the thesis is unchanged, not strengthened.

Xpeng in China has announced Level-4 robotaxi plans built on its own Turing chips, targeting pilots in the second half of 2026 — a second market worth watching, though still at the announcement stage. The application-layer thesis is increasingly being tested in two markets at once, by buyers with structurally similar vision-first approaches.