AI Capex vs. Workforce Contraction | AI Field Notes #1

AnalysisOn April 14, Maine became the first state in the country to pass an 18-month moratorium on new data centers above 20 megawatts. Federal rules only kick in above 100 MW, so Maine is legislating in the gap the White House framework left open. Nassau County, Florida, is now pursuing its own 12-month freeze. The political framing is environmental and ratepayer-driven — water use, grid stress, residential bills — but the real signal is that local authority still has teeth on the one input AI cannot substitute: land attached to power. Trump's Executive Order 14365 is designed to preempt state laws that 'conflict' with federal AI policy; expect the first lawsuit to test whether a moratorium counts. The infra squeeze is not coming, it is here. Compute is abundant. Places to put it are not.

AnalysisOpenAI shipped a major Agents SDK update on April 15 with sandboxed execution — agents can now inspect files, run commands, and edit code inside controlled environments — and announced a mid-2026 sunset for the Assistants API. That second part matters more than the first. Anyone who built on Assistants in the last 18 months gets a migration guide, not a choice. The sandboxing piece is a response to the obvious problem: autonomous agents that touch real systems need real isolation, and 'trust me bro' does not pass procurement. The old assumption was that the Assistants abstraction would be the durable surface. The new one is that long-horizon, tool-using agents are the product, and OpenAI is willing to strand early adopters to make the bet cleanly. Every vendor is converging on the same stack: harness, tools, sandbox, eval. Pick your flavor.

AnalysisTwo months after Trump publicly called Anthropic a 'woke' company run by 'leftwing nut jobs' and the Pentagon blacklisted it, Dario Amodei was in the West Wing on Friday meeting Susie Wiles and Treasury Secretary Scott Bessent. The reason is Mythos — Anthropic's new cybersecurity model that, by its own disclosure, found exploitable vulnerabilities in nearly every major operating system during testing. Suddenly the ideological fight got less interesting to Washington than the access question. The Pentagon still wants unrestricted use for 'all lawful purposes'; Anthropic still refuses mass surveillance and autonomous weapons. Trump told CNBC he had 'no idea' the meeting happened. Classic arsonist-as-firefighter dynamic, with a twist: the lab built a tool scary enough that the administration had to stop boycotting the builder. Safety policy by leverage, not principle.

AnalysisMeta is cutting roughly 10% of its workforce on May 20, with additional rounds planned for the second half of the year, and the framing is already locked in: this is not a contraction, it is an AI reallocation. The same company that spent the last year pitching superintelligence labs and a $14 billion Scale AI deal will now cut about 8,000 jobs while capex keeps climbing. The pattern rhymes with every other layoff this cycle — tech has shed nearly 80,000 jobs in Q1 2026 alone — but the language is new. Middle managers, recruiting, some research functions, all reframed as friction in the path to AI-native efficiency. The interesting question is not whether Meta wanted to cut costs. It is why the word 'AI' now makes quarterly pain sound like strategy.

AnalysisOracle signed a deal on April 13 to procure up to 2.8 gigawatts of Bloom Energy fuel-cell capacity for its AI data centers. Bloom stock jumped 22% the next day. The number matters: 2.8 GW is roughly three nuclear-reactor-units worth of on-site power, deployed as modular fuel cells that can stand up in months rather than the multi-year wait for grid upgrades and substations. Oracle has already raised over $100 billion for its AI build-out. The quiet shift is that hyperscalers are no longer waiting for utilities — they are becoming them. When the grid queue is the bottleneck and state-level moratoriums are landing in Maine, on-site generation is not a hedge, it is the strategy. The next question regulators will have to answer: is an Oracle campus with 2.8 GW of fuel cells still just a customer of the grid, or something else entirely?

AnalysisEY has begun rolling out agentic AI across its entire global assurance division — 130,000 auditors, 160,000 annual engagements, more than 150 countries. The tools are integrated into EY Canvas and are 'expected to support all end-to-end audit activities by 2028.' Framed as augmentation, obviously. The interesting detail is the word 'orchestrate.' When a Big Four firm talks about agents orchestrating complex tasks across their entire professional-services base, the middle rung of the audit career ladder is what is being orchestrated. The classic three-years-of-grunt-work path to senior associate gets compressed by software that does the ticking and tying faster than a second-year. The other three firms are doing the same; Deloitte and KPMG have parallel rollouts. Credential inflation on the way up, automation on the way in. The partnership economics have not caught up yet. They will.

AnalysisSiemens, startup Humanoid, and Nvidia announced on April 16 that the HMND 01 Alpha wheeled humanoid has been running autonomous tote-handling — picking, transporting, placing containers alongside human workers — in Siemens' own electronics factory in Erlangen, Germany. It is not a demo reel; it is a line deployment. The robot runs on Nvidia's physical-AI stack and Humanoid's KinetIQ framework. The money quote from the industry is that 2026 'production is fully committed,' with fleets heading to Hyundai and Google DeepMind. The middle-rung erosion story in logistics used to be about AMRs moving pallets; now it is bipedal-style machines doing the same dexterous hand-work that required a person. MIT Tech Review's piece on gig workers in Nigeria and India strapping iPhones to their heads to record training data is the other half of this story. Humans are training the things replacing them, at piecework rates.

AnalysisDeepSeek V4 appears to have shipped this week — a 1-trillion-parameter MoE with a 1-million-token context, claimed 81% on SWE-bench Verified, and pricing at $0.30 per million tokens. Weights expected under Apache. For context: that is roughly an order of magnitude cheaper than comparable closed frontier models, on a benchmark that went from novelty to saturation inside a year. V4 was reportedly trained in part on Huawei silicon, which makes the export-control debate newly awkward. The pattern is familiar: China ships open weights; OpenAI and Anthropic ship revenue. Both can be true. What shifts today is the default for anyone prototyping a coding agent — the 'free and good enough' option now has benchmark numbers that are hard to wave away, even if independent evals come in a bit lower than the claims.

AnalysisGoogle rolled out side-by-side AI Mode in Chrome on April 16. Click a search result and the webpage opens next to a live AI conversation instead of replacing it. You can drag in open tabs, images and PDFs as context. It is small, and that is the point — Gemini now has 750 million users, and Google is embedding the model where people already are rather than asking them to go somewhere else. Quiet deploys beat loud demos, again. Perplexity has 45 million users and better agent ergonomics; Google has the default browser on a billion-plus devices. The old assumption was that the 'AI browser' would be a new product category. The new reality is that Chrome is the AI browser, and the answer pane sits where the sidebar used to.

AnalysisStanford's 2026 AI Index, released April 14, puts the US performance lead over China at 2.7 percent. In 2023 that gap was somewhere between 17 and 31 percentage points depending on the benchmark. SWE-bench Verified went from 60% to near 100% in a single year. Organizational adoption hit 88%. Four in five university students now use generative AI. The US still vastly outspends — $286 billion in private AI investment versus China's — but benchmark parity while capex diverges is the pattern to watch. Switzerland ranks first in AI talent; the UAE is being called a leading global AI hub. The thing nobody is saying out loud: the dollar-to-capability ratio is getting worse for US labs every quarter, and the export-control moat is not generating the performance gap it was supposed to buy.

AnalysisGeneral Compute announced on April 18 an ASIC-first inference cloud targeted specifically at AI-agent workloads, with general availability set for May 15. The pitch is narrow: agents generate long chains of small, tool-using inference calls, not big one-shot completions, and the GPU stack was designed for the latter. Custom silicon tuned for the former claims big cost-per-call improvements. It is the first credible non-Nvidia inference story aimed at the specific shape of agent traffic rather than training or chat. Nvidia-backed SiFive hit a $3.65 billion valuation last week on a related open-chip thesis. The pattern beneath the announcements: agent traffic is different enough from chat traffic that the infrastructure layer is splitting, and whoever owns the agent-specific cost curve gets a durable margin against the generalist cloud. The hyperscalers will respond. The specialists got here first.

AnalysisA ChinaTechNews piece on April 18, tracking SCMP reporting the day before, puts roughly 80% of global open-source AI developer usage on Chinese-origin models — Qwen, GLM, DeepSeek, Kimi, MiniMax and others. Alibaba's Qwen alone accounts for more than half of open-weight downloads. Zhipu's GLM-5.1 tops SWE-Bench Pro under an MIT license. Meta's Muse Spark, released April 8, is explicitly not open-weight, ending the Llama era without replacement. Put together: the country the US built export controls to slow down now owns the default infrastructure layer for anyone who wants to run AI without paying a US API. The old argument — that Chinese open weights were a curiosity for researchers — does not survive the 80% number. The new shape of the market is a paid US frontier and a free Chinese substrate. Distribution beats model quality. Again.

AnalysisThe Pentagon formally blacklisted Anthropic earlier this year after the company refused to drop safety restrictions on mass surveillance and autonomous weapons. A RedState piece on April 18 notes the obvious follow-up: federal agencies are using Claude anyway, routed through AWS Bedrock, GCP, and cleared contractors, without waiting for the official policy to resolve. The contradiction is not subtle. Agency CIOs have deadlines; the inter-agency standoff does not. This is how AI procurement actually works inside government right now — procurement-by-workaround, with the formal blacklist functioning as paperwork rather than policy. It also sharpens why Amodei flew to Washington this week: the blacklist was costing Anthropic DoD revenue but not much else, and the administration's leverage is narrower than the press releases suggest. The line between cleared and uncleared vendor is getting dotted.