Capacity Is the Product

If You Only Read One Thing

The scarce asset today is not software; it is permission to use capacity. The FCC turned old satellite interference rules into broadband supply, while Anthropic's SpaceX deal shows frontier labs buying whoever has idle power, even a rival. Infrastructure is becoming a market only when rules and politics let someone sell it.

The FCC Sells Capacity

The fastest way to build a new broadband network this week was not to launch a satellite. It was to change the interference rule that told satellites how hard they were allowed to talk.

The FCC on April 30 voted to replace the 1990s-era Equivalent Power Flux Density framework for non-geostationary satellites with performance-based protection criteria for geostationary systems. In the agency's telling, the change could create more than $2 billion in economic benefits and up to seven times more capacity for space-based broadband services, particularly in rural and remote areas. TechRadar framed the consumer impact more plainly: Starlink, Amazon's satellite project, and other low-Earth-orbit operators can run denser service if they can coordinate interference protection under the new regime.

Why it matters: This is a spectrum story, but the mechanism is closer to a zoning change. The physical satellites already exist or are planned; what changed is the regulatory ceiling on how much useful capacity they can deliver in the same bands. Old EPFD limits protected incumbent geostationary operators by assuming 1990s satellite designs, which meant the scarce resource was not only spectrum but permission to use modern radios at modern density.

That shifts the competitive terrain. Satellite broadband has usually been treated as a rural fallback, not a serious substitute for cable or fixed wireless. Moving from static protection to negotiated coexistence makes orbital broadband a more direct rival to terrestrial networks. Starlink benefits first, but the playbook also matters for Amazon Leo and direct-to-device systems such as AST SpaceMobile.

The more important structural point is that capacity in network industries often hides inside rules. Fiber needs rights-of-way. Wireless needs licensed bands. Satellites need interference assumptions. When the rule changes, the supply curve can move before a single new factory or launch pad comes online.

Room for disagreement: The incumbent critique is not frivolous. Viasat and other geostationary operators have warned that looser protections could raise interference risk, and the FCC is relying heavily on private coordination to prevent a public-good problem. If the gains mostly accrue to Starlink, a technical modernization becomes regulatory capture.

What to watch: The test is whether Starlink, Amazon Leo, or another NGSO operator files concrete service, pricing, or deployment changes tied to the new FCC framework by the end of September. If the order only produces lobbying claims, it is less a capacity increase than a deregulatory press release.

Anthropic Buys Rival Power

Elon Musk spent months attacking Anthropic's politics. Anthropic just bought all the spare capacity at one of his most important AI data centers.

Anthropic said it signed an agreement with SpaceX to use all compute capacity at Colossus 1, giving it more than 300 megawatts of new capacity and over 220,000 Nvidia GPUs within the month. The company immediately doubled Claude Code's five-hour rate limits for Pro, Max, Team, and seat-based Enterprise plans, removed peak-hour reductions for Pro and Max Claude Code users, and raised Opus API limits. xAI says Colossus 1 includes H100, H200, and GB200 accelerators; TechCrunch described the result as xAI turning from model company into something that looks like a neocloud.

Why it matters: Yesterday's Anthropic-Google backlog story was about concentration. Today's development is about liquidity. Compute is becoming valuable enough that a model company with weak demand for its own product can sell capacity to a rival with stronger demand.

That is not how hyperscalers usually behave. Google, Microsoft, Amazon, and Meta tend to hoard scarce compute because internal model progress, product distribution, and cloud revenue reinforce one another. xAI's decision to rent Colossus 1 tells a different story: if Grok's demand curve does not justify the hardware, the rational move is to turn GPUs into external revenue before SpaceX heads toward an IPO. Anthropic gets immediate relief from rate limits; SpaceX gets a credible external customer for the orbital-compute story it wants investors to believe.

The lesson is not that compute is the new oil. That phrase hides the important difference. Oil is fungible. AI compute is tied to power, interconnect, cooling, software stack, customer trust, and political permissions. The market becomes liquid only when a buyer can absorb the capacity quickly and a seller is willing to admit that internal demand is not enough.

Room for disagreement: This could be a one-off balance-sheet event rather than a new market. Musk may have overbuilt Colossus 1 while shifting xAI training to Colossus 2, and Anthropic may be unusual because its developer demand is visibly constrained. Most frontier labs will still prefer dedicated capacity over spot-market dependence on a rival.

What to watch: The confirming evidence will be in SpaceX's IPO documents. If external AI compute revenue or customer concentration appears as a named business line, Colossus is no longer just xAI infrastructure; it is a wholesale capacity product.

The Contrarian Take

Everyone says: AI compute scarcity means the biggest labs will vertically integrate, hoard GPUs, and squeeze everyone else out.

Here's why that's incomplete: Scarcity also creates trade when demand curves differ. Anthropic can use Colossus 1 immediately; xAI apparently cannot, at least not at full economic value. The FCC story rhymes with that: capacity did not appear from new hardware, but from a rule change that let existing and planned systems operate closer to their technical potential. The durable winners may not be the firms that own the most infrastructure. They may be the firms that can convert dormant capacity into contracted demand before the financing window asks harder questions.

Under the Radar

Chrome made local AI feel less local - Chrome's reported 4GB Gemini Nano download is a platform-consent problem disguised as a privacy feature. Local inference sounds user-protective, but if the browser quietly allocates disk, bandwidth, and future feature surface without a clear front-door choice, "on-device" becomes another way to shift platform costs onto users. (Source)
The PC hobbyist market is becoming residual demand - PC Gamer reported that Taiwanese motherboard makers are cutting 2026 expectations as memory and CPU shortages hit DIY builders. The structural point is not that gaming PCs are doomed; it is that consumer hardware is now bidding after AI servers for components, so the old upgrade cycle becomes a leftover-capacity business. (Source)

Quick Takes

Kalshi priced the exchange, not the app. Kalshi raised a $1 billion Series F at a $22 billion valuation, double its valuation from five months ago, and says annualized revenue exceeds $1.5 billion. The bet is that regulated prediction markets become exchange infrastructure, with liquidity and compliance replacing novelty as the moat. (Source)
Bumble is killing its own interface. Bumble CEO Whitney Wolfe Herd says the company will say goodbye to swiping after paid users fell 21% to 3.2 million in Q1. The swipe was the category's growth hack and its commoditizer; removing it is an admission that dating apps need differentiated matching quality, not more inventory. (Source)
India has a private launch-company unicorn. Skyroot raised $60 million at a $1.1 billion pre-money valuation as it prepares Vikram-1, the first orbital launch attempt by an Indian private company. India is not just subsidizing space prestige; it is trying to build commercial launch capacity around ISRO's infrastructure and lower manufacturing costs. (Source)
Fitbit became a subscription wedge. Google's $100 screenless Fitbit Air pairs cheap hardware with the rebranded Google Health app and a Gemini-powered Health Coach for Premium subscribers. The device is less important than the bundle: a low-friction sensor that makes an AI health subscription feel continuous rather than occasional. (Source)

The Thread

Today's stories are about the same bottleneck wearing different costumes. The FCC found capacity inside a stale interference regime. Anthropic found capacity inside a rival's underused data center. Bumble is searching for matching capacity after the swipe exhausted itself. Kalshi is turning attention into market liquidity. The common move is rebundling: take an asset that looked fixed, then change the rules, customer, or interface that determines how much of it can actually be used.

Predictions

New predictions:

I predict: By 2026-09-30, Kalshi will announce at least one embedded-distribution partnership with a brokerage, financial-data terminal, sportsbook, or major media platform that routes event contracts outside Kalshi's own app. (Confidence: medium; Check by: 2026-09-30)
I predict: By 2026-08-31, Bumble will make a non-swipe matching surface the default or opt-out default for at least one paying-user cohort, not just describe it as future strategy. (Confidence: medium; Check by: 2026-08-31)

Generated: 2026-05-07 17:05 EDT