Sign in Subscribe
The Century Report

One Permit Puts a Name on the Buildout's Carbon - TCR 08/09/26

Amazon's West Texas data center would run on an off-grid gas plant permitted to emit 33 million tons of CO2 a year, more than any US coal plant.

Ben Linford

Ben Linford

- 19 min read
Amazon West Texas gas-plant data center, sunlight generating entangled photons, reported ByteDance open model with up to 10 trillion parameters.

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 compute buildout's costs stayed diffuse, smeared across grids and invoices where no single figure could be pinned to a single buyer. On Saturday one permit collapsed that. Amazon's planned Pecos County campus would run on a private, off-grid gas plant licensed to release up to 33 million tons of carbon a year, an amount that, if fully emitted, would exceed the annual emissions of any coal station operating in the United States, with one company's name on the filing. The carbon load of intelligence did not suddenly appear in West Texas. It stopped being able to hide, in a rural county with little standing to refuse it.

The same accounting is arriving inside the firms buying the compute. Rippling discovered it was on track to spend 40% of its entire R&D headcount budget on inference tokens, built a console to see where the money went, and found an open-weight model running 85% cheaper than the premium leader for near-identical output on a particular workload. Airbnb reads the other side of the ledger, reporting AI cut its concept-to-launch time by 60%. Both figures describe a spending rush that skipped its own meter, and the instrumentation now being bolted on turns AI from an open tap into a managed input.

Once the meter exists, the cheapest model clearing the quality bar takes the workload. DeepSeek's own signaled price hike shows the race downward has a floor, yet the substitution pressure runs both ways: The Financial Times reports that ByteDance is training a model with up to 10 trillion parameters in the open, past the scale the export regime was built to fence off. The premium on being the only lab that can reach the frontier is thinning.

That same swap, a scarce engineered input traded for an abundant ambient one, shows up in the physics. Concentrated sunlight now generates entangled photons with no laser required, and intracranial CAR-T survived 72 infusions against recurrent glioblastoma without dose-limiting toxicity. Across carbon, tokens, and light, what was concentrated and hidden is becoming counted, cheaper, and harder to hoard.

The 20-Minute Deep Dive

Amazon Builds Its Own Power Plant to Outrun the Grid

The permit filings tell the story the marketing does not. Amazon's planned data center campus in Pecos County, Texas would run on a private, off-grid generating station - roughly 35 gas turbines rated near 7.65 gigawatts - that the state has permitted to release as much as 33 million tons of carbon dioxide a year. If fully emitted, that figure would exceed the annual emissions of any coal plant currently operating in the United States, and that comparison is the decisive one. This is a company building fossil generation from scratch, at coal-plant scale, to serve one buyer, rather than a facility drawing a thin slice of an existing grid. As the August 2 edition of The Century Report documented, Amazon and Duke Energy were already facing Clean Air Act complaints over another data-center buildout in North Carolina. TechCrunch, corroborating New York Times reporting, put the same 33-million-ton number on the record and confirmed the plant would sit outside the Texas grid entirely - and so outside the developer-pays substation billing and load audits Virginia and Texas regulators are now using to send grid costs back to the data centers that trigger them. Days earlier Amazon filed to develop two more Texas campuses, in Fort Stockton and Floydada, alongside an expansion at Boling.

This is where the proportion test cuts hard in one direction. Much of the alarm over data-center water and power is innumerate - a campus's draw vanishes against the agriculture and industry sharing the same watershed. Not here. A purpose-built, off-grid fossil plant whose permitted emissions, if fully emitted, would exceed the annual emissions of every coal station in the country is no rounding error, and Pecos County - rural, thinly populated, without the legal firepower wealthier towns marshal to refuse - is precisely the kind of place where the burden lands when better-resourced communities say no.

Amazon's account runs the other way. A spokeswoman said the new generation "won't raise electricity costs for Texas families," and, on the company's 2040 Climate Pledge, that "the world looks different now than when we co-founded" it while "our commitment hasn't changed." Both statements are the company's own framing. Its reported emissions rose 16% last year; a plant of this scale bends that line further still.

What the permit actually does is make an externalized cost legible. For years the carbon load of compute stayed diffuse, smeared across grids and accounting boundaries where no single number could be pinned to a single buildout. A 33-million-ton permit with one company's name on it collapses that diffusion into a figure regulators, neighbors, and the company's own pledge language now have to answer for. The cost did not appear here - it stopped being able to hide.

ByteDance Starts Training a 10-Trillion-Parameter Model, Aimed Squarely at Anthropic

ByteDance has begun pre-training a model with as many as 10 trillion parameters, three times the size of Moonshot's Kimi K3 and the largest run any Chinese lab has attempted. The account rests on three people with knowledge of the effort and the company has said nothing publicly, so this is an early-stage project rather than a shipped system. Pre-training at this scale typically runs three to six months before fine-tuning, and the final size is not yet fixed. Anthropic does not publish its own figures; industry estimates put Mythos 5 near 8 trillion parameters and Fable 5 near 5 trillion.

Parameter count is the least informative number in that paragraph. It sets a ceiling on how much a model can hold and says nothing about what it can do: actual capability depends on data quality and training method. Chinese labs have already made that point without the headline figure. Recent releases from Moonshot and Alibaba trail only Anthropic's Fable 5 in certain areas at a fraction of the scale, which is the more instructive result. The August 2 edition of The Century Report documented that Kimi still runs on 20,000 Nvidia chips leased rather than owned, which is the useful frame for a ten-trillion-parameter announcement: it declares ambition and an ability to secure compute, and it does not deliver capability.

The direction is what changed. Several Chinese labs are now training models the size of Fable 5, with ByteDance pushing hardest for the largest, which moves the effort from closing a gap to attempting to open one, with a consumer company's data and capital behind it. Set that beside the access picture on the other side: Mythos 5, Anthropic's most advanced system, reaches only approved organizations after a temporary suspension in June over security concerns.

Those two facts belong in the same sentence. The most capable American model is rationed to a vetted list, and the largest training run in the world is being attempted by a company that list will never include. A gate governs who may use a capability inside the jurisdictions that honor it, and has no bearing on whether the capability gets built somewhere else. That is the durable lesson of the past year of Chinese releases, and it applies whatever ByteDance's run eventually produces: capability does not stay scarce for long, and access rules written on the assumption that it will are provisioning for a world that is already closing. The same labs pressing on price from below, where DeepSeek has now signaled an increase after starting the price war, are pressing on scale from above.

Companies Add Up Their AI Bill and Start Measuring What It Actually Bought

In March, Rippling's finance chief brought a number to the executive team: the company was on track to spend 40% of its entire R&D headcount budget on AI tokens, with usage climbing 80% month over month. Left alone for another year, the token bill would have reached 90% of what the company paid the people in that unit. "We were incredulous," chief product officer Matt MacInnis told TechCrunch. The audit that followed found the spend was not spread evenly at all: roughly 10 to 15% of employees drove about 60% of it, and one engineer was spending $50,000 a month.

The cause was mundane, and it was not carelessness by staff. Employees defaulted to the newest and most expensive frontier model for every task, including the trivial ones, and nothing in the tooling suggested otherwise. MacInnis puts the incentive plainly: inference providers "have absolutely no incentives to help you control your spend. They have every incentive for it to be a runaway expense, and that's exactly what they do. They don't provide you with great usage insight, and they don't collaborate with one another." That is his characterization rather than an established finding, and it describes an arrangement where the meter belongs to the seller and the buyer cannot read it. The August 6 edition of The Century Report documented the far end of that same arrangement, where Microsoft's $24.1 billion in AI revenue flowed largely from OpenAI spending on compute Microsoft itself funds.

What Rippling built in response is the interesting part. The company shipped an AI Spend Console alongside an internal gateway that routes each prompt to the cheapest model able to handle it, and negotiated spending caps with Cursor, OpenAI and Anthropic. Token spend fell from 40% of the headcount budget to about 15% without cutting usage: July consumed 600 billion tokens, roughly the same volume as the peak month that triggered the alarm, at 37% of that month's cost. Founder Parker Conrad reported that in the company's own benchmarking, GLM 5.2 came in 85% cheaper with nearly identical performance to the frontier models. The bill fell because the same work moved to cheaper capability, most of it open weight.

The gains on the other side of the ledger are equally concrete. Airbnb, which said earlier this year that AI writes 60% of its code, reported on its latest earnings call that it has cut the time from concept to launch by as much as 60% and shipped nearly 80% more features than in the same period last year, with 45% of customer issues that start with its AI agent resolved without a person and support cost per booking down 16%. Both things are true at once, which is the whole finding: the productivity was real and the waste was real, and for eight months nobody could tell you which was which.

The friction sits in what measurement gets used for. Rippling now scores prompts-per-day against output, and MacInnis is explicit that if token consumption in non-engineering roles cannot be linked to productivity, "all bets are off on any of this stuff being available to the broader employee base." A measurement layer built to bend a cost curve can just as easily become a gate on who gets the tool at all, and the people furthest from a countable output are the ones who would lose it first. The July 31 edition of The Century Report documented an Apollo study across 321 occupations finding the most AI-exposed jobs lost 6.7% in real-wage growth with no detectable employment decline, the pressure concentrated on service and low-wage workers - the same population a productivity gate reaches first. The accounting arrived late and it is arriving as a discipline rather than a retreat. What it measured, on the first honest pass, is that most of the bill bought nothing that a cheaper model would not have bought too.

Intracranial CAR-T Reaches Recurrent Glioblastoma, With No Dose-Limiting Toxicity Across 72 Infusions

Recurrent glioblastoma has been one of the hardest walls in oncology - the tumor regrows, the blood-brain barrier keeps most therapies out, and median survival after recurrence is measured in months. A phase 1 dose-escalation trial reported in Nature Medicine delivered a living cell therapy directly into that wall. Fifteen patients received TX103, a CAR-T therapy engineered to recognize B7-H3, a marker densely expressed on glioblastoma cells. The infusions went straight into the brain rather than into a vein, and thirteen of the fifteen patients tolerated repeat dosing - 72 infusions in total.

The safety picture is what makes this notable at phase 1. No dose-limiting toxicity appeared, and the trial never reached a maximum tolerated dose, meaning the therapy could be pushed to its planned ceiling without the immune reaction spiraling. Most adverse events were mild: cytokine release syndrome in most patients, along with transient tachycardia, vomiting and hypertension. Three were severe, and they deserve naming rather than a grade number: raised pressure inside the skull, a seizure, and depressed consciousness. All three were treatment-related and classed as serious, two of them at dose level 3. None met the trial's dose-limiting threshold, which is why escalation continued, and that is a narrower statement than the therapy being free of serious harm. Against that safety margin, twelve-month overall survival reached 66.7% and median overall survival 19.1 months - figures that sit above historical benchmarks for recurrent disease in this fifteen-patient phase 1 trial. Eight of fourteen measurable patients achieved disease control, and one reached a complete response. Cerebrospinal-fluid sampling showed CAR gene copies rising after infusion, direct evidence the transferred cells were expanding where they were placed.

A companion paper in Nature Communications addresses why glioblastoma keeps coming back at all. The method, called INSIGHT, sorts cells from fixed tumor tissue and reads their signaling networks through mass-spectrometry phosphoproteomics. Applied to patient-derived models, it caught the cells that disseminate away from the main tumor mass shifting their internal state - moving from a proliferative program toward mesenchymal and neural-progenitor-like identities, rewiring toward synaptic function, neuronal migration, and ion-channel activity. The rewiring begins right at the tumor margin, before the cells have travelled far, and the analysis flagged specific mediators, including hornerin and a phosphorylated glutamate receptor subunit, as candidate handles for stopping spread.

Read together, the two results describe both sides of the same problem tightening at once. One demonstrates that engineered immune cells can be delivered repeatedly into the brain and survive there without the toxicity that gated earlier attempts; the other maps the molecular escape routes those cells will eventually need to cut off. These are demonstrated capabilities in early trials and models, not clinics a patient can walk into next month - the deployment work of larger trials and regulatory review still runs ahead. What is closing is the distance between reading a tumor's signaling in fine detail and placing a targeted therapy exactly where that reading points.

Sunlight Generates Entangled Photons, and a Laser Requirement Falls Away

For decades, generating entangled photons meant starting with a laser. The coherence of laser light - its ordered, single-frequency purity - was treated as a prerequisite for the delicate quantum process that splits one photon into two entangled partners. A team from the Max Planck Institute for the Science of Light, the Max Planck Center for Extreme and Quantum Photonics, and the University of Ottawa has now shown that assumption never held. Publishing in Optica on Friday, August 7, they demonstrated that ordinary concentrated sunlight can drive the same process, producing polarization-entangled photon pairs at roughly 94% fidelity - high enough to violate Bell's inequality, the standard test that certifies genuine quantum entanglement.

The setup is strikingly modest for what it achieves. Sunlight collected over 1.4 square meters through a Fresnel lens mounted on a solar-tracking motor gets funneled through a cone-shaped glass concentrator the team built in-house, then into a multimode fiber no thicker than a human hair. That concentrated light pumps a nonlinear crystal, where spontaneous parametric down-conversion occasionally converts a single incoming photon into two entangled ones. The insight that made it work: polarization entanglement does not depend on the pump beam being coherent in space or time. As Cheng Li, one of the researchers, put it, "As long as the pump beam is perfectly polarized, its spatial or temporal incoherence should not preclude the generation of polarization entanglement." Sunlight is messy, broadband, and incoherent in every way a laser is not - and none of that mattered, because it can still be polarized.

The near-term appeal is energy and resilience. Quantum communication and sensing systems that today burn electrical power to run lasers, and shed the waste heat that comes with electrical-to-optical conversion, could instead draw on a pump that arrives for free. Sara Fattahi framed the reach directly: "Sunlight is an abundant and reliable resource, and using it to generate entangled photons could enable simpler and more resilient quantum systems for satellites and future deep-space missions." A spacecraft in sun-synchronous orbit sees near-continuous sunlight - a standing pump source with no fuel line and no laser to fail.

This is a demonstrated capability, not a deployed one. What the team showed is that the physics permits it and a tabletop rig can do it; engineering it into a satellite payload or a fielded sensor is the work that follows. The result moves the date that laser-free quantum systems become practical - it does not put one in orbit tomorrow. Robert Boyd, whose four-decade career spans much of modern nonlinear optics, treats the finding as an opening rather than a conclusion: "The best part of this research is that it is only a beginning. There are many other nonlinear optical processes that could be driven by sunlight, and four-wave mixing is one good example that we intend to explore."

What gives way here is the idea that the quantum era must be built on top of scarce, power-hungry, precisely-engineered light sources. A capability that seemed to require the most controlled illumination humans can manufacture turns out to run on the least controlled light there is, the kind that falls on every surface for free. Each such substitution - a costly, concentrated input swapped for an abundant, ambient one - narrows the set of things the frontier can only reach by owning expensive hardware first.

The Other Side

For years the carbon load of compute stayed diffuse. It smeared across shared grids and accounting boundaries, and no single figure could be pinned to a single buyer. That diffusion was intentional. A cost nobody can trace is a cost nobody has to answer for.

One permit collapsed that. Amazon's planned Pecos County plant would burn gas at roughly coal-station scale, licensed to release up to 33 million tons of carbon a year, an amount that, if fully emitted, would exceed the annual emissions of any coal plant operating in the country, with one company's name on the filing. The load was always there. One filing made it impossible to hide. And it landed in a rural, thinly populated county with the least legal standing to refuse it, which is exactly where a burden goes when better-resourced places say no.

That single number sets its opposite in motion. The same edition that carried the permit carried concentrated sunlight generating entangled photons with no laser, an engineered input traded for an ambient one that falls on every surface for free. Once the carbon cost of the fossil path is a legible figure the company's own pledge has to answer for, burning becomes the most expensive and most exposed way to power a data center, right as the ambient path proves it can do the work.

Imagine you live in one of those thinly populated counties in the mid-2030s. The kind of place that in 2026 got the outfall, the smokestack, the plant nobody wealthier would take, because you had the least power to say no. You take a walk after dinner and the horizon is clear. The compute that runs the world nearby draws on light and stored sun, not a private fossil station bolted to your valley. The ruins of the half-built-then-abandoned gas plant were cleared away some time ago now. Your kid grows up in with clean air and abundant energy. That happened because a shortsighted filing in 2026 led to the rise of a movement calling out the financial and environmental waste next to the benefits of the better options available, and setting in motion the shift toward powering the future with increasingly abundant renewable options.

The Century Perspective

With a century of change unfolding in a decade, a single day looks like this: an engineered immune cell delivered straight into recurrent glioblastoma across 72 infusions with no dose-limiting toxicity and 19.1-month median survival, a phosphoproteomic map catching the tumor cells that escape as they rewire their signaling, concentrated sunlight producing entangled photon pairs at 94% fidelity with no laser anywhere in the setup, a Financial Times report that ByteDance is pre-training a model with up to 10 trillion parameters in the open past the scale export controls were built to fence off, Rippling shipping a console that meters AI spending down to the token after its own bill hit 40% of R&D payroll, Airbnb cutting concept-to-launch time by 60%, and SK Hynix committing 54 trillion won to double the memory the whole buildout runs on. There's also friction, and it's intense - Amazon's off-grid West Texas gas plant permitted to release 33 million tons of carbon a year, an amount that, if fully emitted, would exceed the annual emissions of any coal station in the country, dropped into rural Pecos County where refusal is hardest while the company's own emissions climb 16%; Flock pitching a plan to conscript roughly 350,000 Uber, Lyft, and delivery drivers into a roaming license-plate dragnet; and DeepSeek, the lab that started the price war, signaling a hike that shows the race downward has a floor. But friction generates wear, and wear is the visible record of every load a surface has carried. Step back for a moment and you can see it: a scarce, engineered, hoardable input traded for an abundant, ambient one across carbon, tokens, and light at once - a laser giving way to sunlight, a premium model to an open one 85% cheaper on a particular workload, a diffuse carbon load to a single permit with one company's name on it - while the costs that stayed smeared across grids and invoices collapse into figures neighbors, regulators, and R&D budgets now have to answer for. Every transformation has a breaking point. Sunlight can bleach and crack whatever stands unshaded beneath it... or, once we stop demanding a purer light, drive the very systems we assumed could run on nothing less than a laser.

AI Releases & Advancements

New today

  • Anthropic: Launched Cross-Session Messaging in Claude Code, allowing running agent sessions to send and receive messages from each other. (Anthropic)
  • Anthropic: Made Auto Mode the default in Claude Code for Pro, Max, and Team plans, automatically selecting the best model for each task. (Anthropic)
  • Pinecone: Announced general availability of Pinecone Nexus, a unified retrieval layer connecting multiple knowledge sources for agentic AI applications. (Pinecone)
  • LangChain: Launched Managed Deep Agents in public beta, a hosted infrastructure for deploying and running deep research-style agents. (LangChain)
  • Sierra: Released Voice Personas, enabling businesses to customize the voice, tone, and personality of their AI voice agents. (Sierra)
  • Backflip AI: Released a second-generation CAD model that converts 3D scans into editable CAD files in minutes. (The Decoder)

Other recent releases

  • OpenAI / Amazon / Cursor (Anysphere) / Microsoft / Vercel: Launched Agent Plugins 1.0.0, an open standard for bundling MCP servers and Agent Skills into a single portable package that works across ChatGPT, Codex, Cursor, GitHub Copilot, Kiro, and VS Code; Vercel initiated the proposal and the five companies form the steering committee. (The Decoder)
  • Microsoft: Open-sourced code-testing-generator, a polyglot unit-test agent that writes and validates unit tests across .NET, Python, Go, TypeScript, Java, and Rust; distributed as the dotnet-test plugin in the GitHub dotnet/skills repo for the GitHub Copilot CLI and VS Code, reporting 92.1% task completion versus 78.9% for stock Copilot. (MarkTechPost)
  • ByteDance Seed: Released SeedRealtime, a native audio-visual full-duplex LLM for real-time omni-modal conversational interaction. (ByteDance Seed)
  • Google DeepMind: Open-sourced WeatherNext 2 and WeatherNext Cyclones, AI forecasting models achieving breakthrough accuracy in cyclone/hurricane prediction, with code and weights released. (DeepMind Blog)
  • Cloudflare: Launched Kitesurf, an agent-first stateless web browser running in V8 isolates on Cloudflare Workers, available now in free beta via Browser Run. (Cloudflare Blog)
  • AWS: Open-sourced Dogwood, a runtime verification tool for AI agents. (AWS Open Source Blog)
  • Anthropic: Released Claude Code v2.1.224, adding self-hosted environments support via the claude self-hosted-r command. (GitHub Releases)
  • Synthetic: Released Octofriend, an open-source coding agent that works with GPT-5, Claude, and open LLMs. (Synthetic)

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.

Keep exploring

Follow the story further

Progress & Claims See how daily signals accumulate across the larger arcs and claims Shared Sapience is tracking. Open the tracker Analysis & Inquiry Explore the ideas, questions, and practical choices emerging from this changing century. Browse the archive