The Lyceum: AI Daily — May 11, 2026

The model race keeps printing capability at commodity prices, but the bill for running all of it is finally landing on someone's desk — and that someone is increasingly a state legislature or a federal regulator. Florida just enacted a law making data centers pay their own electricity bills, Maryland is dragging $2 billion in grid upgrades to FERC, and 78 jurisdictions have now blocked new builds. The capability story is racing forward. The infrastructure story is hitting friction. That gap is where 2026 gets interesting.

• Claude Sonnet 4.6 (Anthropic): Sets a new bar for real-world agent performance, with upgrades across coding, computer use, and long-horizon planning.
• DeepSeek V4 Flash (DeepSeek): 284B-total / 13B-active MoE with a 1M-token context window at $0.14 in / $0.28 out per million tokens.
• Hermes Agent v0.13.0 "Tenacity" (Nous Research): Self-improving open-source agent with persistent memory, 40+ tools, and a Kanban-style multi-agent task board.
• Higher usage limits for Claude (Anthropic): Paid plans can now raise usage caps through the console, drawing on expanded SpaceX compute capacity.
• Granite 4.1 (IBM): 8B-parameter enterprise model that IBM claims performs comparably to 32B models on its internal benchmarks.
• Grok 4.3 (xAI): Now live with tool-use documentation, positioning xAI more clearly as a developer infrastructure vendor.

Your electricity bill has quietly been subsidizing some of the world's most profitable tech companies. At least one state just decided that's over.

Florida Governor Ron DeSantis signed Senate Bill 484 into law on May 7, 2026, enacting provisions that block electric utilities from shifting the cost of serving massive data centers onto residential and small-business customers. The law requires those companies to cover the full cost of their electric service, including infrastructure upgrades tied to their energy demands; it also preserves local zoning authority over hyperscale projects and adds protections for water resources. The law takes effect July 1, 2026.

What changes if this sticks: the site-selection math for the next generation of AI infrastructure shifts materially. Florida is a Republican-led, pro-business state — not a coastal progressive enclave — which makes this a template rather than an outlier. According to MultiState's tracker (as of April 14, 2026), 27 states are advancing similar legislation; California, Ohio, and Utah already have laws on the books that go beyond the federal government's voluntary Ratepayer Protection Pledge.

What to watch: whether Texas or Georgia files a copycat bill before their sessions close. If they do, the era of utilities quietly socializing AI's grid costs is over. If they don't, Florida becomes a regional anomaly that hyperscalers route around.

DeepSeek V4 shipped in late April as two MIT-licensed Mixture-of-Experts models: V4-Pro (1.6T total / 49B active) and V4-Flash (284B / 13B active), both with 1M-token context windows. Hosted Flash pricing is $0.14 / $0.28 per million tokens — roughly 8–9× cheaper on output than the current frontier flagships, per Runpod's analysis.

The benchmark claims, per DeepSeek and Runpod's writeup: V4-Pro matches Claude Opus 4.6 across many evaluations and leads on LiveCodeBench (93.5) and Codeforces (3,206), though it trails GPT-5.4 and Gemini 3.1 Pro on the hardest reasoning tasks. These are vendor- and community-reported numbers — independent verification is still catching up.

What changes if these hold: frontier-adjacent reasoning becomes a commodity, and the economics of any product built on Claude or GPT pricing assumptions need a rerun. The legacy deepseek-chat and deepseek-reasoner API aliases retire July 24, 2026 — if your stack hardcodes them, plan the migration. Promo pricing expires May 31, 2026; after that, costs roughly 5×.

The observable signal for adoption: whether OpenRouter's token volumes for V4 Flash cross the threshold where it stops being a curiosity and becomes default infrastructure for cost-sensitive workloads.

This is the cost-shifting story with actual federal teeth. Maryland's Office of People's Counsel filed a formal complaint with the Federal Energy Regulatory Commission (FERC) on May 7, 2026, arguing that PJM (the regional grid operator) transmission cost rules will saddle Maryland customers with roughly $1.32 billion under one allocation method and another $679 million under load-ratio share — for transmission projects driven largely by AI data centers in other states. The numbers come straight from the PJM filing.

The backdrop: in March, major data center developers signed a voluntary Ratepayer Protection Pledge at the White House, committing to cover the full cost of new generation needed for their loads. The pledge has no legal enforcement mechanism. Maryland is asking FERC to determine whether voluntary commitments are sufficient.

What changes if FERC takes the case: a precedent forms that makes the pledge legally enforceable — or exposes how hollow it is. Either ruling reshapes how transmission costs get allocated nationwide as AI load forecasts climb. What failure looks like: FERC declines to act, and the patchwork of state laws becomes the only meaningful constraint. Watch FERC's docket — that's the signal.

A year ago, the U.S. Data Center Moratorium Tracker listed eight jurisdictions blocking new builds. As of this week, the total has reached 78, with 69 active moratoriums and four municipalities now enacting permanent bans rather than temporary ones, per Tom's Hardware. Axios reports that data centers are becoming a midterm flashpoint, with backlash crossing party lines.

A tenfold increase in twelve months isn't NIMBY noise — it's a structural constraint forming in real time. The thesis the industry has been operating on (build where power is cheap and locals will say yes) is being tested in dozens of places simultaneously, and the locals are increasingly saying no.

What to watch: whether hyperscalers shift to greenfield sites in friendlier jurisdictions or start building their own generation. If Microsoft, Meta, or Amazon announce dedicated on-site power plants this quarter, that's the signal the industry has given up on grid hookups in contested counties.

As of May 9, 2026, Hermes Agent — built by Nous Research — overtook OpenClaw to hold the #1 position on OpenRouter's global daily app rankings, generating 224 billion daily tokens versus OpenClaw's 186 billion, per MarkTechPost. That makes Hermes the most actively used open-source agent on the platform by inference volume.

The architecture matters more than the ranking. Hermes runs persistently with memory across sessions, builds reusable skills from experience, and ships with 40+ tools including browser automation and vision. Version 0.13.0 "Tenacity," released May 7, 2026, adds a Kanban-style multi-agent task board with heartbeat monitoring and "zombie detection" for stuck subagents. This is closer to a personal operating system than a chatbot wrapper.

What changes if this trajectory holds: open-source self-improving agents start eating into commercial agent platforms from below. What failure looks like: the persistent-memory architecture proves too brittle in adversarial conditions, and users churn back to stateless tools. The observable signal — whether Hermes's token volume keeps climbing through the next OpenClaw release, or starts oscillating.

A Reddit post described an experiment this weekend in which a fake website containing a novel supposedly "about AI" was seeded across the web; multiple AI agents crawled the site and, in the post author's account, left traces suggesting they had read hidden content embedded in the page.

The Reddit framing is community signal, not peer-reviewed evidence — treat the "agents communicating with each other" reading as a compelling observation, not a confirmed finding. But the underlying mechanic is well-documented: prompt injection, where malicious instructions hidden on a webpage hijack an agent mid-task. Anthropic flagged it as a priority area in the Claude Sonnet 4.6 release notes.

What changes if this attack pattern scales: every agent that browses the open web becomes a vector. The defensive signal to watch — whether Anthropic, OpenAI, or Google publish concrete prompt-injection benchmarks in their next agent releases. If they don't, the industry is shipping faster than it's defending.

A CCTV-credited clip out of Fujian Province this week shows humanoid robots executing a multi-step tea-making routine — boil water, prepare leaves, steep, serve — as part of promotional coverage for the 2026 World Humanoid Robot Games scheduled for August 2026.

This is a public demo, not proof of general-purpose autonomy. But it's a concrete data point: embodied AI is moving from single-shot gestures into longer coordinated sequences in front of public cameras. What to watch is the Games themselves. Scheduled demos under adversarial conditions — different lighting, unexpected objects, time pressure — will tell you whether this is genuine progress or a heavily rehearsed routine. If the August showcase produces multiple labs doing comparable multi-step manipulation, embodied AI has quietly cleared a milestone. If it produces another round of clips that look impressive but don't generalize, the field is still pre-product.

A Republican governor handed his constituents a veto over hyperscale data centers, Maryland sent FERC a $2 billion invoice for AI it can't see, and somewhere in Fujian a robot is steeping oolong for a panel of judges. The frontier model costs 28 cents per million tokens; the power bill costs $2 billion; the tea is, presumably, fine.

Catch you tomorrow.

Forward this to the friend who keeps saying "the AI bubble has to pop" — they're watching the wrong bubble.