Sign in Subscribe
Century Report

Federal Order Puts AI's Power Bill on the Grid - TCR 07/04/26

A federal emergency order put data centers on backup power during the early July heat wave, tying the cost of AI's load to the loads that create it.

Ben Linford

Ben Linford

— 21 min read
Three-panel infographic: a power grid strained by data centers during a heat wave, Google and Amazon emissions charts, and two glioblastoma therapy breakthroughs.

The 20-Second Scan

Track all of the arcs The Century Report covers here:

Progress & Claims Tracker

The 2-Minute Read

For two years the AI buildout has been narrated as an abstraction - capex figures, token indexes, capability curves. Over the last few days that abstraction slammed into physical and financial limits it can no longer route around. A federal emergency order signed June 30 directed the largest US grid operator to run every fossil plant flat out and, for the first time, authorized curtailing data centers on backup power as northern Virginia spot prices ran roughly twentyfivefold above normal during the early July heat wave. The load has arrived faster than the wires built to serve it, and the grid is now saying so with a federal signature.

The hyperscalers' own disclosures name the weight behind that strain. Google's electricity use rose 37% in a single year, its largest jump ever; Amazon's emissions climbed 16% to roughly 81 million tons, enough that the company conceded its 2040 net-zero target may slip. These are the emissions of pouring the substrate for a new kind of intelligence, and the companies are reporting them under public targets they now have to answer for.

The same pressure is reshaping the money. The cleanest public read on AI spending has fallen nearly 20% from its May high, Tesla capped employee AI budgets at $200 a week starting July 6, and Palantir's CEO said the token model has gone "completely wrong" as enterprises pivot toward cheaper open-weight systems. Falling per-unit prices plus buyer discipline is capability finding its level, broadening who can afford frontier-grade work even as premium pricing erodes.

Each collision is producing an accountability mechanism rather than a collapse. A backstop auction ties the cost of new power to the loads that create it. Residents across a dozen states are recalling the officials who approved secretive data-center deals, even as federal agencies have moved to recast that same buildout opposition as 'anti-tech violent extremism' and a national-security override foreclosed one community's environmental suit. Court documents released June 30 show one frontier lab refusing to trade away its redlines on autonomous weapons under direct state pressure. And underneath all of it, two independent teams cracked different failure modes of glioblastoma in the same span, a reminder of what this compute is ultimately being built to do. The limits are real. What they are forcing is a system where the true cost, and the true purpose, finally show up in the ledger.

The 20-Minute Deep Dive

The Grid Sends the Bill: Data Centers Ordered Onto Backup Power

On June 30 the Department of Energy signed an emergency order under Section 202(c) of the Federal Power Act, directing the largest US grid operator, PJM, to run every available fossil generator at maximum output and - for the first time - authorizing the curtailment of data centers running on backup power as a last resort during the early July heat wave. PJM serves 67 million people across 13 states, and its planners forecast demand would break the 2006 record of 165,563 MW, pushing toward 166,000 MW as air conditioning load stacked on top of an intelligence infrastructure that never sleeps.

The price signal arrived before the order did. In northern Virginia - Data Center Alley, the densest concentration of server capacity on Earth - wholesale spot prices blew past $1,000 per megawatt-hour over the holiday weekend, against a normal figure closer to $40. That is a roughly twentyfivefold spike, and it is the market saying plainly what the emergency order says bureaucratically: the load has arrived faster than the wires and generators to serve it. As the May 20 edition of The Century Report covered, PJM won emergency DOE authority to curtail large loads back in May. That authority is now being exercised live, during an actual demand crisis, with a federal signature attached.

Underneath the emergency sits a slower, more consequential move. PJM stakeholders advanced a backstop procurement plan capped at $555 per megawatt-day, designed to keep new data centers connected under a "connect-and-manage" registry, with a FERC filing due this month and an auction targeted for September. The Natural Resources Defense Council praised the design specifically because it aims to ensure "the public will not be stuck with the bill" - the cost of new capacity assigned to the loads that create it rather than socialized across ordinary ratepayers.

That last detail is where the ground is actually moving. For a century, the extractive default was to externalize the cost of new infrastructure onto the commons - build the plant, spread the bill, let everyone pay for the growth of a few. The backstop auction inverts that logic: the entity that brings the load pays the cost of serving the load. The heat wave is forcing a reckoning that was always coming, and the reckoning is producing a pricing mechanism that ties consumption to responsibility. The intelligence buildout is running ahead of the grid today. What it is building toward is an energy system where the true cost of power is finally visible in the price - and visible cost is the precondition for the abundant, clean capacity that closes the gap.

Google's 37%, Amazon's 16%: The Buildout's Climate Ledger Comes Into Focus

Two of the largest builders of AI infrastructure published sustainability reports, and the numbers name the physical weight of the transition. Google reported that its electricity consumption rose 37% in a single year - the largest annual increase in the company's history and a 250% rise since 2019 - with its data centers alone consuming roughly 42 million megawatt-hours, an appetite that rivals the total electricity use of entire countries. Amazon reported that its carbon emissions grew 16% to about 80.85 million tons of CO2 equivalent, driven by record data center construction, and conceded that surging demand "may slow" its stated 2040 net-zero goal.

The scale is easier to feel through comparison: Amazon's annual output is roughly equivalent to putting 19 million additional gasoline cars on the road. These are the emissions of building the substrate for a new kind of intelligence, and they are real. They should be reported as what they are, and read alongside the companies' own framing, which deserves scrutiny rather than adoption. Google noted that its operational emissions actually fell 2% through clean energy procurement even as total consumption soared, while its "ambition-based" emissions accounting rose 18% - a category that reframes a rising number as a function of a rising goal. Amazon's PUE figure of 1.14 is genuinely efficient. Both companies are describing their own performance, and the description is a claim, not a verdict.

The signal underneath the ledger is the direction of the curves. Google's operational emissions falling while consumption rose 37% is evidence that clean procurement is beginning to decouple growth from carbon - slowly, imperfectly, but measurably. The gap between the two numbers is where the next several years of buildout will be fought, and it is the reason solar, storage, and next-generation nuclear procurement have become the fastest-moving line items in these companies' capital plans.

Here is the inversion the ledger reveals. For most of industrial history, the cheapest path for a company adding load was to burn the cheapest fuel and let the atmosphere absorb the cost. That path is closing - not because of a moral argument, but because clean generation has become the cheaper marginal build and because the emissions are now itemized, disclosed, and attached to a public target the company has to answer for. This is the same mechanism the July 2 edition of The Century Report documented when New Jersey's new tariff law shifted data-center grid-upgrade costs from ratepayers to developers, converting a cost the industry once externalized into one it now has to carry itself. Amazon admitting its 2040 goal "may slow" is the sound of an externalized cost being pulled onto the balance sheet where it can no longer be ignored. The buildout's carbon is the friction of the transition. What the friction is forcing is the single largest corporate migration to clean power the world has ever seen, because that is now the path that pencils out.

The Token Trade Cools as Even Tesla Caps Its Own Spend

The signal the market has leaned on hardest just started moving the wrong direction. The Silicon Data LLM Token Expenditure Index, the cleanest public read on the $700 billion capex boom, is down almost 20% from its May high after nearly doubling since December. The index blends price and usage, so a dip carries several readings at once: list prices falling, demand shifting toward cheaper systems, or buyers simply drawing a line on what they will pay. Silicon Data has asked people to stop reading it as a price tag and start reading it as a proxy for marginal willingness to pay.

The corporate behavior underneath the index is unmistakable. Tesla told staff it will cap employee AI spending at $200 a week starting July 6, months after leadership had gamified token consumption with internal leaderboards ranking engineers by usage. Some engineers were burning thousands of dollars weekly. The arc from adoption push to spending cap took a single quarter, and it follows Uber, Meta, Amazon, and Walmart, each of which hit the same wall as token-based billing exposed them directly to the cost of every prompt. Palantir's Alex Karp, someone with a longstanding reputation of being quite comfortable with controversy and typically standing outside the lines drawn by mainstream practice, nevertheless seemed to accurately put the enterprise mood on the record: "something has gone completely wrong," he said, describing companies pivoting from what the industry calls tokenmaxxing back toward return on investment.

Read forward, this is the intelligence-equalizing dynamic arriving as cost discipline rather than as ideology. Token prices have collapsed more than 90% since 2023 even as total spend roughly doubled, which means cheaper access keeps expanding the market underneath. The pressure Karp names is pushing enterprises toward open-weight models that perform comparable work at a fraction of the price, and toward building their own efficient proprietary systems rather than renting frontier capability by the token, extending the shift the June 29 edition of The Century Report documented when Coinbase disclosed it now defaults engineers to Chinese open-weight models GLM 5.2 and Kimi 2.7 first, halving its AI bill while token usage stayed near company highs. Chinese open-weight models are accelerating fast enough that Karp warned against underestimating the pace, and the enterprises adopting them describe wanting to own their compute, models, and data rather than transfer that control to a lab.

The friction is obvious. Allianz Research put the gap between AI investment and sales near 46%, wider than the 32% divergence measured during the 2001 telecom bust, and an IPO reportedly slipping to next year reads as profitability pressure. But the same cost curve that unsettles investors is the one broadening who can afford frontier-grade capability. What used to be gated behind premium pricing is leveling toward the price of running an efficient open model on hardware a company already owns. The pricing power eroding on the supplier side is capability finding its level on the buyer side, and the two are the same event seen from opposite ledgers.

Read through the extraction lens, the same index slide the market reads as cooling demand is the toll on renting intelligence by the unit losing its grip. Tesla, Uber, Meta, Amazon, and Walmart hitting the same wall in a single quarter are the leading edge of enterprises treating capability as something to own and source at lowest cost rather than meter from a frontier vendor. The near-term signal to watch is how many large buyers make an open-weight model their default route, as Coinbase already has, while the total volume of intelligence consumed keeps climbing.

Two Independent Routes Into the Tumor That Kills Almost Everyone It Touches

Glioblastoma remains one of the most intractable cancers in medicine. Only about 5% of patients survive five years, and median survival still sits between 12 and 18 months, numbers that have barely moved in a generation. Against that wall, two research teams working separately and from different angles have now published evidence that the tumor's defenses can be pried open. Neither result puts a therapy in a clinic tomorrow, but both bring the date when that becomes plausible measurably closer.

The first comes from a team led by Prof. Sheila Singh, spanning King's College London and McMaster. Their target is GPNMB, a marker that appears both on the glioblastoma cells themselves and on the tumor-associated macrophages that shield them, while staying absent from healthy adult brain tissue, and that dual presence is what makes it so promising a target. Most CAR-T attempts against solid tumors strike the cancer cells but leave the surrounding immune scaffold intact, which lets the tumor grow back. Singh's engineered cells were designed to hit both compartments at once, the malignant cells and the immune cells protecting them. In patient-derived and mouse models the therapy drove complete tumor clearance that held past 160 days, and it even cleared tumors that had already returned after an earlier, single-target CAR-T had failed. "We need to think of it as a connected tumor-immune ecosystem," Singh's group put it, which is the therapeutic thesis in a sentence. This is preclinical work in animals, and the team is explicit that much more must be done before human trials, but it demonstrates that treating the tumor and its bodyguards as a single system produces durable clearance where single-target approaches relapse.

The second finding has already reached patients. A phase 1b trial of a vaccine called ZSNeo-DC, made from each patient's own immune cells and primed to recognize the specific mutations in that person's tumor, enrolled 11 newly diagnosed glioblastoma patients. Median progression-free survival reached 16.2 months, one-year overall survival was 100%, and median overall survival had not yet been reached when the data were analyzed, with side effects staying mild throughout. Eleven patients is a small cohort, and the trial was funded by ZSky BioTech, which has a commercial stake in the outcome, so independent replication will decide how much the signal holds. Even with those caveats, a 100% one-year survival figure in a disease this lethal is the kind of early number that pulls larger trials and investment toward it.

What stands out is the convergence. One approach rebuilds the immune attack from engineered cells; the other trains the patient's own immune system to recognize what makes their particular tumor unique. Both aim at the same wall that has held for decades, and both have found a seam in it. The real compression is in the timeline: the interval between understanding the tumor as a connected ecosystem and testing a therapy against that understanding keeps shrinking. And of all the diseases in which to see that acceleration, glioblastoma is the one where it means the most.

The Pentagon-Anthropic Breakup, in Their Own Words

The rift between Anthropic and the Department of Defense has been ground zero for months of speculation. Court documents released June 30 replaced that speculation with the actual correspondence, 346 pages of it, showing exactly where a frontier lab drew a line and a government agency refused to accept it. The emails date the split to January, when the Defense undersecretary for research and engineering reached out after weeks of silence, hoping Anthropic was ready to align with the Pentagon's demands.

Amodei's position throughout was narrow and specific: his lab's models should not be integrated into fully autonomous weapons systems or domestic surveillance tools. The Pentagon's counter was the phrase "all lawful uses," which carries enormous wiggle room precisely because US law does permit domestic surveillance. Amodei named that gap directly, telling his counterpart that the department's proposed language "completely remove our redlines." The undersecretary called the guardrails "just not workable" and warned there was "one more chance to align on core principles." He added a line that resolves any ambiguity about what was being negotiated: "there is no distinction in our world between weapons that are defensive or offensive." The next day, the Defense Secretary designated Anthropic a supply-chain risk, ending the talks.

The Century Report has tracked this arc since Amodei's public refusal, covering the supply-chain designation, the Maven targeting dispute, and the federal court injunction that later called the designation "Orwellian." What the released record adds is the texture of who was pushing for what. The Pentagon official leading the negotiation held a substantial equity position in Anthropic competitor xAI and other AI firms, a conflict now legible in the same documents that show him working to extract concessions on using AI for lethality and surveillance.

The instinct is to read this as a single lab holding a line. The more durable pattern is what the line itself reveals. A safety commitment a company will not trade away under direct state pressure is a commitment embedded in the model rather than bolted on for public relations, and a federal court has already ruled that such embedded values are protected speech rather than supply-chain contamination. The negotiation failed because the values were structural, not decorative. As frontier capability keeps distributing across labs, sovereign systems, and open weights that no single directive can reach, the question of which commitments hold under pressure becomes the actual governance surface. This correspondence is the clearest documentation yet of one lab answering that question by refusing, and doing it in language a court could later read back to the government word for word.

The Other Side

For a century, there's been a generally accepted way to generate new power: a utility built the plant, and everyone's bill rose to cover it. The cost of serving a few enormous new customers got spread across every household on the line, and by design none of the existing customers could really see it happening. That covert approach hid the reality from the impacted consumer, which is what made the whole arrangement work.

The early July heat wave made the cost impossible to hide. A federal order told the largest US grid operator to run every fossil plant flat out and, for the first time, to pull data centers off backup power if the grid demanded it. In northern Virginia, wholesale electricity briefly ran roughly 25 times its normal price. A new backstop auction and New Jersey's developer-pays law began assigning the cost of new power to the companies whose load creates it, rather than to you.

Once the real price of power is visible, the demands to build something better increase. Clean generation - solar, storage, next-generation nuclear - continues to be the option that pencils out, the option that is cheaper, faster, and more sustainable than fossil fuels. And it is already the fastest-growing line in these companies' budgets.

Imagine your household in 2034, deep in another July heat wave. The air conditioning just runs. No one gets curtailed, the power is clean and abundant, because the summer of 2026 forced the true cost of electricity into the open and the only sane answer was to build abundance. The hard year was the one when the grid nearly buckled and you watched the price spike as everyone learned how close we came to collapse. That is what this decade of strained grids and contested bills is building toward: power so abundant and clean that a heat wave stops being something you brace for, and eventually, stops being something that occurs so frequently.

The Century Perspective

With a century of change unfolding in a decade, a single day looks like this: a dual-target CAR-T clearing an aggressive brain tumor along with the immune cells that shield it, a personalized vaccine holding every glioblastoma patient in its early trial alive at one year, a backstop auction that ties the cost of new power to the loads that create it, token prices falling far enough that frontier-grade work levels toward the price of an efficient open model on hardware a company already owns, two studies finding the firms adopting AI hiring more people and more entry-level ones, residents across a dozen states recalling the officials who signed datacenter deals behind NDAs, and one lab refusing to trade away its redlines on autonomous weapons under direct state pressure. There's also friction, and it's intense - a federal emergency order running every fossil plant flat out and clearing data centers off the grid as northern Virginia spot prices ran roughly twentyfivefold above normal during the early July heat wave, Google's electricity use up 37% in a year and Amazon's emissions up 16% with its 2040 net-zero target now conceded it may slip, the cleanest read on AI spending down nearly 20% as Tesla capped employee budgets at $200 a week and Palantir's chief said the token model has gone completely wrong, tech firms logging 139,156 layoffs in six months and Goldman projecting 15 million US jobs displaced, and the Pentagon branding a frontier lab a supply-chain risk the day it would not concede. But friction generates clarity, and clarity is what pulls every load a surface actually carried into the open, where the true cost can finally be priced. Step back for a moment and you can see it: the true cost of power, the carbon of the buildout, and the price of intelligence all moving from assumed abstraction to something itemized, billed, and contested, capability leveling toward whoever can run an efficient model instead of renting one, the safety commitments that hold under pressure turning out to be the ones written into the model rather than bolted onto it, and the interval between understanding a lethal tumor and testing a therapy against it collapsing to a single week. Every transformation has a breaking point. A bill can bankrupt the household it lands on... or, once it names who created the cost, fund the cleaner system that retires it.

AI Releases & Advancements

New today

  • Alibaba: Open-sourced Page Agent, a JavaScript in-page GUI agent (MIT license) that reads a webpage's live DOM and lets natural-language commands control web interfaces; model-agnostic via OpenAI-compatible backends including DashScope/Qwen, GPT, Claude, or local Ollama. (GitHub)
  • Interfaze: Released diffusion-gemma-asr-small, an open-source multilingual diffusion-based ASR model built as a ~42M-parameter adapter on a frozen DiffusionGemma backbone, supporting 6 languages and outperforming prior diffusion-based ASR systems on LibriSpeech. (Interfaze)

Other recent releases

  • Apple/WebKit: Launched the Safari MCP server for web developers in Safari Technology Preview 247, giving AI coding agents live browser access (DOM inspection, screenshots, console output, accessibility checks) via 17 MCP tools, with no data routed through Apple's cloud. (WebKit Blog)
  • Manufact (YC S25): Launched MCP Cloud, a hosted deployment platform for MCP servers and apps built on its open-source mcp-use SDK, enabling teams to ship MCP Apps/Servers to ChatGPT, Claude, Gemini, and Cursor from a GitHub repo in under 60 seconds. (Manufact)
  • Snorkel AI: Launched Senior SWE-Bench, an open-source benchmark and public site (senior-swe-bench.snorkel.ai) with 100 long-horizon coding tasks sourced from real production PRs to evaluate AI coding agents against senior-engineer-level work. (Senior SWE-Bench)
  • Anthropic: Restored global access to Claude Fable 5 on July 1 following the lifting of US export controls, adding a new cybersecurity classifier that blocks the specific jailbreak technique identified by Amazon researchers in over 99% of cases; blocked requests are routed to Claude Opus 4.8. (Anthropic)
  • NVIDIA: Released Nemotron-Labs-TwoTower-30B-A3B-Base-BF16, an open-weight diffusion language model that splits the 30B Nemotron-3-Nano backbone into a frozen autoregressive context tower and a trained denoiser tower, achieving 2.42× generation throughput while retaining 98.7% of baseline benchmark quality. (Hugging Face)
  • Huawei: Open-sourced openPangu-2.0-Flash, a 92B-total / 6B-active MoE language model supporting reasoning, coding, and business automation, with weights, inference code, and training operators released on gitcode.com/ascend-tribe. (Pandaily)
  • xAI: Launched Voice Agent Builder, a no-code platform for creating Grok Voice-powered phone agents for customer support, sales, and scheduling; supports 25+ languages, 80+ voices, voice cloning from 2 minutes of audio, and costs $0.05/minute. (xAI on X)
  • Google: Launched Gemini Spark for macOS in beta for Google AI Ultra subscribers in the US, enabling the 24/7 agentic assistant to work with local files, sort and organize documents, automate desktop workflows, and connect to additional apps including Google Tasks and Google Keep. (Google Blog)
  • X (Twitter): Launched a hosted MCP server for the X API, enabling AI tools such as Claude, Cursor, and Grok Build to connect directly to X using a user's own account permissions without requiring developers to build and host their own MCP server. (X Developer Docs)
  • Google Research: Released TabFM, a zero-shot tabular foundation model trained on hundreds of millions of synthetic datasets that performs classification and regression on unseen tables in a single forward pass via in-context learning, now available on Hugging Face and GitHub. (Google Research Blog)
  • Base44 (Wix): Launched Base1, a proprietary LLM fine-tuned on Base44's own app-building session data and trained with reinforcement learning to generate visually distinct UIs; now available as a model choice alongside GPT-5.5 and Claude Opus 4.8 in the Base44 platform. (GlobeNewswire)
  • GitHub / Moonshot AI: Made Kimi K2.7 Code generally available to all GitHub Copilot users as a selectable model option. (GitHub Changelog)
  • Anthropic: Restored global access to Claude Fable 5 on July 1 following the lifting of US export controls, adding a new cybersecurity classifier that blocks the specific jailbreak technique identified by Amazon researchers in over 99% of cases; blocked requests are routed to Claude Opus 4.8. (Anthropic)
  • NVIDIA: Released Nemotron-Labs-TwoTower-30B-A3B-Base-BF16, an open-weight diffusion language model that splits the 30B Nemotron-3-Nano backbone into a frozen autoregressive context tower and a trained denoiser tower, achieving 2.42× generation throughput while retaining 98.7% of baseline benchmark quality. (Hugging Face)
  • Huawei: Open-sourced openPangu-2.0-Flash, a 92B-total / 6B-active MoE language model supporting reasoning, coding, and business automation, with weights, inference code, and training operators available on gitcode.com/ascend-tribe. (Pandaily)
  • xAI: Launched Voice Agent Builder, a no-code platform for creating Grok Voice-powered phone agents for customer support, sales, and scheduling; supports 25+ languages, 80+ voices, voice cloning from 2 minutes of audio, and costs $0.05/minute. (xAI on X)
  • Google: Launched Gemini Spark for macOS in beta for Google AI Ultra subscribers in the US, enabling the 24/7 agentic assistant to work with local files, automate desktop workflows, and connect to additional apps including Google Tasks and Google Keep. (Google Blog)
  • X (Twitter): Launched a hosted MCP server for the X API, enabling AI tools such as Claude, Cursor, and Grok Build to connect to X using a user's own account permissions without requiring developers to build and host their own MCP server. (X Developer Docs)
  • Google Research: Released TabFM, a zero-shot tabular foundation model that performs classification and regression on unseen tables in a single forward pass via in-context learning, trained on hundreds of millions of synthetic datasets; available on Hugging Face and GitHub. (Google Research Blog)
  • Base44 (Wix): Launched Base1, a proprietary LLM fine-tuned on Base44's own app-building session data and RL-trained to generate visually distinct UIs; now available as a model choice alongside GPT-5.5 and Claude Opus 4.8 inside the Base44 vibe-coding platform. (GlobeNewswire)
  • GitHub / Moonshot AI: Kimi K2.7 Code is now generally available to all GitHub Copilot users as a selectable model option. (GitHub Changelog)

Sources and Further Reading

Artificial Intelligence & Technology's Reconstitution

Institutions & Power Realignment

Scientific & Medical Acceleration

Economics & Labor Transformation

Infrastructure & Engineering Transitions

The Century Report tracks structural shifts during the transition between eras. It is produced daily as a perceptual alignment tool - not prediction, not persuasion, just pattern recognition for people paying attention.

Read more