Signal

What it is: the Anthropic Economic Index measures how AI is being used across the economy — which jobs, which tasks, automation vs augmentation — from real usage data. The connector makes it directly queryable ("which occupations use AI most?", "what do teachers automate?"). Until now this was a dataset for researchers; now it's a tool.

Why it matters to you: it's the closest thing to ground truth on the AI-jobs question, from the company with the data — a primary-source counterpart to academic studies like Stanford's. Also a quiet strategic tell: labs publishing labour-impact data proactively are building the evidence base for the policy fight ahead.

Anthropic's own data about Anthropic's own product — selection bias is structural. Cross-read against Stanford, not instead of it.

The hard numbers: $140M Series B closed in February. First commercial factory at Morgan Hill, California — deliberately next to engineering HQ, the same playbook Tesla used with Fremont in 2010. Production starts second half of 2027, ramping toward 40 GW of capacity, roughly 10,000 of its "Heron Link" units.

What the product actually is: not a traditional transformer. It collapses several pieces of mechanical power-conversion equipment into one modular, software-controlled unit for data centres, solar and battery projects — built to get them connected to the grid years faster than the current transformer queue allows, and to ride out large voltage swings.

Why it matters to you: transformers are the least glamorous item on every data centre's critical path and routinely the longest lead time. Whoever breaks that bottleneck sells to every hyperscaler at once. Baglino's Tesla pedigree is what makes this credible rather than another power startup.

Honest caveats: still developing its manufacturing process; needs "more real runtime" at test sites before banks treat it as financeable. Watch: does the factory actually open on time in 2H 2027 (it's in the claims ledger).

The core argument: Anthropic and OpenAI can monetise compute better than anyone else, so they outbid everyone — cloud providers, enterprises, nations — until they control most of the world's usable FLOPs within a few years. Economies of scale in training, compute scarcity, and eventually recursive self-improvement all push the same direction; the unresolved question in the episode is whether anything pushes back.

The number that should stop you: over $10 trillion of total AI capex by the end of the decade — with a serious discussion of whether hyperscaler debt at that scale raises interest rates globally, drives non-AI-exposed countries toward bankruptcy, and crashes non-AI equities. A sovereign debt crisis as an AI side effect.

Also in there: labs are shifting compute from serving customers (inference) back into R&D — a tell about how close they think self-improvement is. And China: gets under 10% of new compute, but its labs need less to stay in the race.

Patel talks his book like everyone — SemiAnalysis sells this worldview. But his track record on chips is the best there is, which is exactly why he opens the claims ledger below.

What was said: a Micron fellow told Hot Chips that the wafer-cost penalty HBM (the stacked memory every AI accelerator depends on) carries against ordinary DDR5 grows with each generation. Each new HBM generation eats more silicon area for the same capacity — the opposite of how semiconductor economics is supposed to work.

Why this is a weak signal worth watching: everyone models AI costs as "chips get cheaper every year." If the memory half of the bill structurally doesn't, the cost curves in every AI buildout model are wrong, and memory makers (Micron, SK Hynix, Samsung) hold more pricing power than the market assumes. Samsung's answer, announced the same day, is to move the computing into the memory itself — see consensus below.

One source, one conference talk. But it's an engineer speaking against his own industry's marketing, at the industry's own technical conference — the opposite of hype.

The tally so far: a minority investment in Cloverleaf Infrastructure (powered-land development, announced last week); $1.5 billion into SB Energy, which builds "power-first" AI infrastructure; a stake in Lancium (Blackstone-backed, integrates grid interconnection + behind-the-meter generation + storage); anchor investor in KKR's Helix Digital Infrastructure; plus NVentures positions across nuclear, battery storage and grid software over two years.

The read: the company selling the shovels has concluded the mine's real constraint is electricity, and is vertically integrating into it. When the best-informed player in AI hardware spends billions securing power rather than fab capacity, that's the market telling you where the scarcity is. Earnings Wednesday.

Why this belongs in your brief: it's the second state-actor story this cycle where the lab, not a government, was the detecting party. The labs are becoming chokepoints states must route around — which is also the strongest argument governments will use for deeper access to them. Watch how this shapes AI regulation arguments over the next months; "we catch influence ops" is becoming the labs' national-security bargaining chip.