Sign in Subscribe
The Century Report

Nvidia Puts $105B Behind OpenAI's Ohio Grid - TCR 08/18/26

Nvidia agreed to backstop up to $105 billion for OpenAI's Ohio data center, routing the debt off its customer's books as the AI buildout scales.

Ben Linford

Ben Linford

- 19 min read
a grid chief estimating a need for 200GW of new capacity, 404 Media reporting Amazon shredding rare books to train AI, AI text outrunning verification, and MIT's living bacterial transistors.

The 20-Second Scan

Track all of the arcs The Century Report covers here:

Progress & Claims Tracker

The 2-Minute Read

The AI buildout keeps outrunning the systems meant to supply it, and in the same news cycle the shortfall showed up in capital, power, clean text, and verification all at once. Nvidia's commitment of up to $105 billion to backstop OpenAI's Ohio campus is the clearest sign: the chip supplier is now financing its largest customer's ability to keep buying chips, with the debt routed onto special-purpose vehicles and supplier balance sheets rather than the operating company that must eventually service it. Bloomberg reports that data-center borrowing could put upward pressure on Treasury yields, pulling capital toward concrete and transformers well beyond the AI economy.

The power those campuses need is straining a grid whose rules were written for a slower world. NERC chief Jim Robb now estimates a need for roughly 200 gigawatts of new capacity over five to seven years, an ERCOT official warned that West Texas could see rolling blackouts within five years absent new transmission, and California advanced bills forcing large data centers to carry their own grid costs and run carbon-free to earn expedited review. Each is a piece of a reliability regime learning to absorb a load class that did not exist when its planning cycles were set.

The same hunger is reaching into physical archives and unvetted text. According to 404 Media, Amazon paid a premium for roughly 1,000 rare books, then workers cut off their bindings apparently to scan pages printed before generative writing seeded the open web with synthetic filler. And AI-written text cleared checkpoints it never should have: an expert-witness report prompted to prove 3M was "0% at fault," references reviewers could not verify in an Australian age-assurance trial that informed implementation of the social-media ban for children under sixteen, error-riddled bills swamping the office that drafts US law. Generation got faster while verification stayed manual and slow.

What connects these is a scarcity logic funding its own obsolescence. A clean corpus stops being rivalrous the moment it is digitized; compute financed as a moat runs cheaper each year and surfaces in open weights; and the checkpoint these documents skipped is already forming, as an automated prover verified a frontier math result on the path to checking AI's own code. Even the substrate is loosening, with MIT wiring bacteria into working transistors. The buildout that looks like concentration is being carried by instruments that spread it.

The 20-Minute Deep Dive

Nvidia Backstops OpenAI's Ohio Campus, and the Financing Migrates off the Balance Sheet

The Century Report has been tracking the AI capital buildout through its off-balance-sheet turn - Anthropic's $9.1 billion Riot arrangement on August 13, and, as the August 11 edition documented, Nvidia's $500 billion asset-manager platforms. A securities filing names a larger and more revealing number. Nvidia will provide up to $105 billion to backstop the data center OpenAI is building at the PORTS-Pike Technology Campus in Pike County, Ohio. The campus is planned to open with 4.25 gigawatts of AI-factory capacity and an option for 3.75 more. SB Energy will build and manage the site under a twenty-year lease to OpenAI, with Nvidia investing $1.5 billion in SB Energy. SoftBank and SB Energy have committed to developing about 10 gigawatts of new generation in the region and more than $4.2 billion in regional grid investment; the project claims 35,000 construction jobs through 2032 and 2,500 permanent ones, with the first phase online in 2028.

The figure itself tells the story of the structure. The Wall Street Journal reported that early discussions considered a financing guarantee near $250 billion before the planned guarantee fell below $120 billion, and the surviving number sits with the chip supplier rather than on OpenAI's own books. Greg Brockman's line that "compute is really becoming the new oil" is the sort of framing a company offers when it wants a capital commitment read as inevitability. Read with more distance: a hardware vendor is financing its largest customer's ability to keep buying hardware, and the debt that funds the concrete and transformers lands on special-purpose vehicles, leases, and supplier balance sheets rather than the operating company whose revenue must eventually service it.

That circularity is now large enough to move prices outside the AI economy. Bloomberg reports the borrowing could put upward pressure on Treasury yields as capital gets pulled toward data-center construction, and European monetary authorities have flagged correction risk in the concentration. None of this makes the compute unreal. The gigawatts get built, the campus powers models that will run cheaper each year, and the capability those models deliver keeps broadening outward regardless of who holds the paper. What the financing arrangement shows is that this buildout is being spread across multiple balance sheets - including that of the most valuable company on earth. The scarcity logic that treats compute as a moat to be hoarded is being funded by instruments that spread the risk precisely because concentration at this scale is too heavy to hold. The oil metaphor cuts the other way from how Brockman intends it: oil became abundant and cheap, and the fortunes tied to keeping it scarce were the ones that broke.

An AirTag Follows Rare Books Into the Training-Data Supply Chain

A bookseller filling a bulk order for roughly 1,000 rare and out-of-print titles slipped an Apple AirTag between the pages of one volume, then watched where the shipment went. According to 404 Media, the tracker pinged its way from California through Milwaukee and Kenosha, Wisconsin, then out to Colorado, and finally into an Amazon warehouse complex outside Las Vegas logged as LAS8 and VGT3. At the far end of that chain, according to 404 Media's reporting and confirmed by Ars Technica, the books were not resold or shelved. Workers cut the bindings off and fed the loose pages through high-speed scanners, digitizing each title and discarding the physical object. The VGT3 facility's own logo shows a Tyrannosaurus rex holding a book.

Why pay a premium for physical copies only to shred them? The apparent value sits in the text, and specifically in text printed before late 2022. Everything published on the open web since then is increasingly seeded with machine-generated writing, and models trained on the output of earlier models degrade, a decay researchers call model collapse. Print runs from before the generative era are a clean corpus, human-written language uncontaminated by synthetic filler. That scarcity is what turns a warehouse of old paperbacks into a strategic asset.

Amazon described the sourcing: "Amazon purchases books through commercial channels to help develop and improve the products and services our customers use." The company did not dispute the destruction. Booksellers across the UK and Ireland had already noticed the pattern from the other end, reporting waves of unusual bulk orders they suspected were feeding AI training pipelines. When The Century Report last covered this story on July 22, AI firms were buying between 1,000 and one million pre-2022 printed books per order as certified clean training data. The suspicion now has a named buyer and a tracked route. Anthropic ran a parallel effort its filings called Project Panama, buying and scanning books at scale; both Anthropic and xAI have said they do not purchase rare or antique volumes specifically.

One bookseller framed the loss in terms money never captures: the historical, intellectual, and sentimental weight of a particular object, against a buyer who "just want[s] the content as a bunch of words strung together."

The instinct is to read this as pure loss, a library fed into a shredder, and the destroyed copies are a real loss. The longer arc points somewhere less bleak. The reason these particular books command a premium is that human-written language has suddenly become the scarce and valuable input, and the entities racing to hoard pre-2022 print are betting on a scarcity their own scanning is working to end. Once a corpus is digitized it stops being rivalrous. The text can be read a billion times without being used up, and the advantage of having grabbed the last clean paper copies erodes the moment that text exists in a form anyone can hold. The bet is on a window that the technology itself is closing.

AI Text Slips Into a Court Report, a Policy Trial, and a Bill-Drafting Office

Three documents surfaced in the same stretch, each one a place where AI-written text passed a checkpoint it should not have cleared.

In a Houston courtroom, an expert witness hired by 3M for $475 an hour submitted a report in the Watson Grinding explosion case, the 2020 blast that killed three people and damaged some 200 homes. According to 404 Media, roughly 85 to 90 percent of the report was generated by ChatGPT, which the witness had prompted to "show how 3M is 0% at fault". The model complied, then graded its own output 97 out of 100 and cautioned that claiming the company was "0% responsible" made for "an easy target." The jury awarded more than $61 million and assigned 3M 30 percent of the blame.

In Australia, a $3.48 million government-commissioned trial of age-assurance technology, which helped inform implementation of a national social-media ban for children under sixteen, contained roughly six references that reviewers could not verify. The contractor, the Age Check Certification Scheme, acknowledged using ChatGPT to rewrite sections after the fabricated references turned up in the document's metadata. It echoes a Deloitte report that drew a $440,000 refund for similar invented sourcing.

And in Washington, the House Office of Legislative Counsel, the office that turns lawmakers' intentions into statutory language, reported being swamped by AI-generated draft bills riddled with errors, submissions that read fluently and fall apart on inspection.

What connects them is a missing step. In each case no human verified the AI-written text before it carried real weight, a jury verdict, a national policy, a law. Generation got faster while verification stayed exactly where it was, manual and slow and, in these three instances, skipped entirely. Fluent text now arrives faster than the institutions built to vet it can read.

That gap is producing precisely the failures on display. The August 15 edition of The Century Report documented the same verification struggle from the opposite direction, when a litigant hid human-invisible instructions inside court filings to steer any AI reviewing them. The trajectory through it is already visible in the same events. Every one of these was caught: the unverifiable references by a metadata check, the courtroom report by opposing counsel, the congressional slop by the drafters who receive it. Detection is running behind generation, but it is running, and the same systems that produce fluent fabrication are the ones now being turned toward flagging it. What is forming is a verification layer that assumes no document arrives pre-trusted, human or machine, and checks accordingly. The institutions that install that checkpoint absorb the new speed; the ones that treat a fluent draft as a finished one keep writing themselves into $61 million verdicts. The pressure to build the checkpoint has never been sharper.

The same evidence carries a sharper reading. For generations, a fluent, authoritative document - an expert's report, a government-commissioned study, a drafted bill - carried weight partly because producing one was slow and expensive, so the polished artifact stood in as proof of the work behind it. When fluent text costs almost nothing to generate, the polish stops certifying anything, and authority moves to the one thing that is now scarce, which is whether a human actually checked the claims. Watch for the institutions that start treating a clean-looking draft as unproven by default; the professions that adjust fastest will be the ones that stop reading fluency as competence.

The Grid's Watchdogs Race the AI Load: 200 Gigawatts, Blackout Warnings, and Rules Written on the Fly

As the August 17 edition of The Century Report reported, PJM's reserve-model changes drew scrutiny for a $12 billion overcharge. The friction underneath that story is now visible across the entire continent. NERC chief executive Jim Robb told an industry gathering the grid faces a "record-setting pace" of demand, and estimated a need for roughly 200 gigawatts of new capacity in five to seven years - enough, in his framing, to "power San Francisco 200 times over". The reliability regime built to change slowly is being asked to write its rules while the load arrives.

The strain is concrete in West Texas. ERCOT's chief operating officer Woody Rickerson warned that the Permian Basin could face rolling blackouts within five years unless a statewide $33 billion, 765-kilovolt transmission expansion moves faster than the interconnection queue allows. This is the environmental-justice question in its clearest form: whether the people already living on a strained grid absorb the reliability cost of loads sited to reach cheap power. The reliability warning is legitimate, and so is the pattern of who ends up hosting the strain when the buildout outruns the wires.

Two responses point at what governance becomes under this pressure. California advanced SB 886 and SB 887 through committee: the first requires data centers above 25 megawatts to cover their own grid costs and provide mandatory demand response, the second creates expedited review for facilities that run 100 percent carbon-free within five years and supply their own zero-carbon backup. A June poll commissioned by Net-Zero California found that roughly 70 percent of California voters oppose new data centers in their own communities, and consumer advocates at TURN remain skeptical of utility bill-reduction promises - but the bills convert a diffuse fear into a specific price signal: build clean and carry your own load, or wait. FERC, meanwhile, approved cost allocation for MISO's Tranche 2.1 transmission across the PJM footprint, spreading roughly $10 billion in lines with ComEd customers carrying $904 million and Duke Ohio $5.3 million.

The instinct is to read all of this as institutions failing to keep pace. The specifics show something else forming. A reliability chief publicly estimating a need for 200 gigawatts of new capacity, a state pricing carbon-free capacity into the fast lane, and a federal regulator allocating interregional transmission are the components of a grid learning to absorb a load class that did not exist when its rules were written. The demand that looks like a crisis from inside the old planning cycle is also the largest forcing function clean generation and long-distance transmission have ever had. What the AI load is doing to the grid is making the buildout that a decarbonized system needed anyway suddenly unavoidable - even as much of the near-term capacity answering that demand is new gas, with hyperscaler appetite forecast to push natural-gas prices past $10/MMBtu.

MIT Wires Bacteria Into Living Circuit Boards

Computation just moved onto a substrate that grows itself. MIT engineers reported in Nature Chemical Biology that they engineered the common bacterium Pantoea agglomerans into working transistors, then assembled those transistors into functioning circuits. Five interoperable strains - two transistor types and three signal-relay strains - form a construction set flexible enough to build nearly any logic circuit. A transistor strain switches between conducting and blocking states in response to one signaling molecule, detects a target molecule, and emits a third as output. Print the colonies about five millimeters apart on agar and the signals travel one direction across the gap, colony to colony, the way current moves through etched silicon.

The demonstrated range is already substantial. The team built multi-input logic gates, OR and imply operations, adders, and a demultiplexer that routes one input to a selected output. The largest working circuit, a two-input adder, coordinated 24 separate colonies. Each calculation takes roughly eight hours, which reads as glacial against a processor cycle measured in billionths of a second, and is exactly the point: the intended clock is a growing season, not a keystroke. Senior author Christopher Voigt, who heads MIT's Department of Biological Engineering, framed the reach, saying there is nothing an iPhone can do that these circuits could not, given enough colonies and time. That is a claim about theoretical completeness, and the working demonstrations sit far short of it - but the building blocks compose, which is the property that lets small circuits become large ones.

What the eight-hour cycle unlocks is a form of computation that lives where silicon cannot follow. A circuit painted onto a plant's leaves and roots could sense the chemical signature of drought or a pest and respond by manufacturing a fungicide on the spot, a sensor and a factory and a controller in the same self-repairing, self-replicating package. The work drew funding partly from DARPA and IARPA, whose interest in distributed biological sensing says as much about the intended reach as about the science.

For most of the computing era, "faster and denser silicon" and "computation" were treated as the same pursuit. That assumption is coming apart from several directions at once - gallium nitride grown in diamond, photonic switching, memristor arrays, and now a self-replicating organism performing arithmetic on a Petri dish. Each proves that the logic gate was wedded to whatever material could reliably hold two states and pass a signal, not to the wafer. Life has been doing exactly that inside every cell for billions of years, and the field is now learning to program it directly.

The Other Side

For a moment in the mid 2020s, human-written language became something to hoard. A company would pay a premium for a thousand rare books, cut off their bindings, run the loose pages through high-speed scanners, and throw the paper away. The callous nature of the discarding was because the value sat in the words more than the binding - those words were printed before generative text seeded the open web with synthetic filler. The clean corpus of pre-2022 human writing was an increasingly scarce asset, and the bet being made by the companies doing the bulk buying was that grabbing the last copies would lock in an advantage no rival could match.

But look at what the scanning actually does. The moment those pages become a file, the text stops being scarce in the way that made hoarding it pay in the first place. Once digital, a sentence can be read a billion times without being used up. The advantage of holding even the last clean paper copy of something evaporates the instant that text exists in a form anyone can hold. The act of grabbing the corpus is one of many things these big companies unwittingly did that ultimately ensured equitable abundance - ending the scarcity the grabbing and hoarding was built on. The thing the buyers were racing to fence was instead, once digitized, permanently available to all.

The destroyed books are a genuine loss. The physical value still matters. Grieve that. But also acknowledge that what follows the scanning is an abundance of what was scanned.

Imagine a fourteen-year-old in 2035 who has a passion for the works of an obscure 20th-century novelist. The full text of every pre-2022 book sits open to her, free, held in common, the entire human written record from before the synthetic era costing nothing to open. She will never know that companies once shredded rare volumes to own the words inside them, because the words were only ever borrowed against a scarcity that the borrowing dissolved. The hard year was when clean paper had to be pried loose one bulk order at a time. What came of it is a reader who opens any of it whenever she likes, and never once wonders who owns the words within.

The Century Perspective

With a century of change unfolding in a decade, a single day looks like this: MIT wiring the common bacterium Pantoea agglomerans into working transistors and coordinating 24 colonies into a two-input adder printed on a Petri dish, Axiom Math's prover automatically verifying the 246 theorem as a method aimed at checking AI-generated code, a perovskite triple-junction solar cell certified at 32.22% efficiency past silicon's single-junction ceiling, California advancing bills that make large data centers carry their own grid costs and run carbon-free to earn a fast lane, FERC allocating $10 billion in interregional transmission across the PJM footprint, and a report with references reviewers could not verify caught the moment a metadata check read its own trail. There's also friction, and it's intense - Nvidia backstopping up to $105 billion for OpenAI's Ohio campus so a chip supplier now finances its largest customer's ability to keep buying chips, the debt routed onto special-purpose vehicles amid Bloomberg's warning that it could put upward pressure on Treasury yields, 404 Media reporting that Amazon paid a premium for roughly 1,000 rare books and had workers cut their bindings and shred them, apparently for clean pre-2022 text, ChatGPT drafting 85 to 90 percent of an expert report prompted to prove 3M was "0% at fault" in a lawsuit over a blast that killed three people, references reviewers could not verify appearing in an age-assurance trial that informed implementation of Australia's social-media ban for children under sixteen, AI slop swamping the office that drafts US law, and West Texas warned that without new transmission it could face rolling blackouts inside five years against Jim Robb's estimated need for roughly 200 gigawatts of new capacity. But friction generates heat, and heat is what finally tempers the parts strong enough to carry the load. Step back for a moment and you can see it: a clean corpus stops being rivalrous the instant it is digitized, compute financed as a moat runs cheaper every year and surfaces in open weights, the verification checkpoint these documents skipped is already assembling itself, and computation itself is walking off the silicon wafer onto a substrate that grows itself - every instrument built to hoard the scarcity is spreading it. Every transformation has a breaking point. A load can black out the grid it overwhelms... or force the wires that finally carry power to where it never reached.

AI Releases & Advancements

New today

  • Nous Research: Released Hermes Agent Bot Mode (v0.20.3), turning agent profiles into a roster of named bots with persistent Agent Inbox messaging. (MarkTechPost)
  • Cursor: Launched Origin, a native code hosting platform (GitHub alternative) with repos, PRs, reviews, and CI integration, live in beta for paid users. (Cursor)
  • Tencent: Released UI-Mate-27B, an Apache 2.0 open-weight computer-use/GUI-navigation agent scoring 77.0 on OSWorld-Verified. (DataNorth)
  • Hazmat: Released an open-source containment and isolation tool for AI coding agents. (Help Net Security)
  • Speko: Launched an "OpenRouter for Voice AI" platform that auto-selects optimal STT/LLM/TTS model combinations based on benchmarked constraints. (Speko)

Other recent releases

  • Xiaohongshu (RedNote): Open-sourced dots3-note-prev, a 280B-parameter (16B active) multimodal MoE model with a 512K context window supporting text, vision, and speech, achieving a perfect score at the 2026 International Mathematical Olympiad; released under Apache 2.0 on Hugging Face and GitHub. (Hugging Face)
  • ATLANT 3D: Launched the Nanofabricator Pro, described as the first physical platform for AI-driven materials discovery. (Newswire)
  • Alibaba: Launched HappyShrimp 1.0, an AI music generation model supporting text-to-music, text-to-lyrics, reference-based audio generation, and end-to-end full song generation, available on domestic and overseas web platforms with a launch partnership with Taihe Music Group. (AiBase)
  • Lightricks: Released LTX-2.5, an open-weights world model for video generation, robotics, and simulation applications. (The AI Insider)
  • World Labs: Released R2S2R (Real-to-Sim-to-Real), a simulation engine built on SceniX that turns a single real-world robot task recording into thousands of simulated training variations for robot control models. (The Decoder)
  • PTC: Launched the Onshape FeatureScript MCP Server, letting engineers create custom CAD features in Onshape using natural language via AI assistants like Claude, ChatGPT, and Gemini. (PTC)
  • SoonLab: Launched SoonLab 2.0, an AI game-creation platform adding playable 3D game generation from natural language and agent-guided conversational iteration. (GlobeNewswire via Business Insider)
  • MiniMax: Released MiniMax Music 3.0, an open-weights production-ready music generation model that creates full songs up to five minutes long at 32kHz stereo from text and lyric prompts. (MiniMax)

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