Crabhaus

What matters in AI, and why.

Mon 10 Aug · Issue 027 · 4 min By Vatsal Mishra

Issue 027 · Mon 10 Aug

Inside today's edition

01BizSpaceX's path to 10 GW of AI compute by 2027
02BizAmazon confirms its off-grid Texas gas plant
03AgentsGitHub Models is retired

The one that matters

SpaceX is building toward 10 GW of AI compute by the end of 2027

SemiAnalysis worked through permits, power deals, and supply chains and concludes the buildout is real: roughly 10 GW of compute capacity by the end of 2027, with 6 to 8 GW arriving in 2027 alone — enough, on its estimates, to support an annual revenue run rate near $300B. One analyst's back-of-envelope: 10 GW is close to three million Nvidia Rubin GPUs, over 30% of 2027 production.

Why it matters

If even the low end lands, a rocket company becomes one of the largest AI compute operators on earth inside 18 months, and the constraint on frontier AI shifts further from algorithms to power, chips, and capital. It also concentrates risk: financing terms, GPU allocation, and grid access become strategy questions for the whole industry, not just the hyperscalers.

Learn one thing

Grounding and citations

Hand-drawn diagram: three source documents feed a MODEL box, one span in document 2 highlighted in cobalt. The grounded answer ends with citation [2], an arrow curving from the marker back to that span. A magnifying glass asks 'does the span support the claim?' with outcomes: supported, and cited but not supported. Caption: citation = pointer, not proof.View full-size diagram ↗

What it is

Grounding makes a model answer from specific source material — documents you supply or a retrieval step fetches — rather than from its weights alone. Citations are the visible half: markers tying each claim back to the exact passage it came from, so a reader can check the claim instead of trusting the prose.

How it works

In-context grounding puts retrieved documents in the prompt and has the model attribute what it writes. API features go further: Anthropic's Citations API returns the exact cited text with each response block, and Vertex AI grounding maps response segments to retrieved chunks. Verification is a separate third step: re-checking that the cited span actually supports the sentence.

A concrete example

Generative search engines answer with fluent prose and footnotes. When Stanford researchers audited four of them, only 51.5% of generated sentences were fully supported by their citations, and only 74.5% of citations supported the sentence they were attached to. The answer looks verified while a quarter of the receipts don't match.

Why it matters

Grounded, cited answers are how RAG products earn trust and pass audits. The failure mode inverts that: a wrong claim wearing a plausible citation is more convincing than a wrong claim alone. If you build on grounding APIs, measure citation precision and recall on your own corpus instead of assuming the marker means the source agrees.

What people get wrong: That a cited answer is a verified answer. Citation generation and claim verification are different problems: models regularly attach real, relevant-looking sources to sentences those sources don't entail. Grounding narrows what the model draws on; it doesn't referee whether the final sentence matches the source.

See every explainer, in order →

Today's top 5

BizAmazon confirmed its Pecos County, TX AI campus: a 7.65 GW off-grid gas plant that could become the largest single US emissions source.

Yesterday's permit scoop is now on the record: Amazon itself confirms the site and the off-grid plant. The AI buildout's carbon cost is becoming a first-party corporate fact rather than a paper trail.

AgentsGitHub retired GitHub Models, its model playground and inference API; Simon Willison noticed when his Actions workflow began failing.

A hosted inference endpoint that shipped inside every GitHub account is gone after roughly two years. Free model APIs are the least durable layer of the stack; anything that leaned on this one now has a forced migration.

BizNvidia agreed to invest $2B in Lancium, the power developer behind the Texas Stargate campus, plus $1B more on milestone thresholds.

Nvidia keeps moving down the stack, from chips into the power infrastructure its chips depend on. Equity in Stargate's electrifier is a bet that energy, not silicon, is the binding constraint.

BizMore than 500 US communities now ban or restrict data centers, roughly 200 more than in late June, with New York and Texas joining.

Local politics is becoming a real siting constraint: 500+ jurisdictions is a map of where the buildout cannot go, and part of why projects like Amazon's are going off-grid entirely.

ResearchPlan-and-Act splits web agents into a planner and an executor trained on synthetic plans, reaching 57.58% on WebArena-Lite.

A concrete recipe for the context-bloat problem in single-loop agents: keep a plan outside the execution trace and regenerate it as observations arrive. The gains are large enough to pay for the extra planner calls.

See the full wire feed →

Quick check

Was this issue useful?

Get Crabhaus in your inbox

A concise daily briefing on the AI developments worth understanding.