Sign in Subscribe
Century Report

Three Companies Pry Open the Chip Layer - TCR 06/25/26

OpenAI built its own AI chip, IBM stacked transistors past the physics wall, and Qualcomm bought the code that frees AI to run on any silicon, all in one cycle.

Ben Linford

Ben Linford

— 18 min read
Three-panel infographic: decentralized compute and energy substrates; surveillance oversight scaling with deployment; AI turning routine EKGs and blood draws into advanced diagnostics.

The 20-Second Scan

Track all of the arcs The Century Report covers here:

Progress & Claims Tracker

The 2-Minute Read

A single day's evidence kept circling one question: who holds capability, and how fast it slips into more hands. OpenAI and Broadcom designed their own inference chip, IBM stacked transistors past the lithography wall, and Qualcomm paid nearly $4 billion for the software that lets AI code run on any silicon. Three companies pried at three separate locks - the vendor, the node, the proprietary software layer - that together kept advanced compute behind a single accelerator stack. The substrate of intelligence is being redrawn from several directions at once, faster than any one actor can close its fist around it.

The same inversion ran through the grid. Sunrun, Tesla, and Renew Home pooled 16.8 gigawatts of capacity out of hardware already bolted to people's walls, pitched directly at the data-center hot spots straining the system, while Base Power put $95 backup batteries into congested PJM territory. Set beside the Department of Energy's $17.5 billion loan toward ten central reactors, two answers to the same demand curve are forming. One assembles power from federal financing over a decade; the other dispatches latent capacity this summer, owned by the households living on it.

Even the friction carried the pattern. AI surveillance contracts doubled to a record $513 million, yet the accountability layer arrived alongside the deployment: Bristol officials killed risk-scoring models they could no longer trust, a mayor's veto forced a procurement process into existence, and advocacy groups turned opaque spending into a public ledger. Washington's thaw toward Anthropic, meanwhile, ran less on any published rulebook than on who proved pleasant in the room - a private accommodation governing a capability that already reproduces on hardware no export order reaches.

Medicine showed the gentlest version of the same diffusion. An AI read sudden-cardiac-death risk from an ordinary EKG, and plasma profiling surfaced a survival signature from a routine blood draw, even pointing toward a dietary intervention. The information was always present in the signal. What changed is that reading it no longer depends on a rare specialist's eye, which is how a capability stops being a privilege and starts becoming a floor.

The 20-Minute Deep Dive

Custom Silicon Multiplies as OpenAI, IBM, and Qualcomm Each Attack a Different Chokepoint

Three moves in one cycle pried at three separate locks holding the compute layer in a few hands. OpenAI and Broadcom announced Jalapeño, an inference chip designed from scratch around the way large language models actually run, built in nine months from OpenAI's own model roadmap and claiming performance-per-watt "substantially better than current state-of-the-art" pending a full technical report. A model company designing the silicon it runs on is the buyer-side answer to the Nvidia vendor lock that has defined this era's economics.

IBM took aim at a different ceiling. Its new "nanostack" architecture fits nearly 100 billion transistors on a fingernail-sized chip, roughly twice its prior density, by delivering the compute gains a sub-1-nanometer node would imply without building features physically smaller than a nanometer, which the laws of physics forbid. Director Jay Gambetta framed it as a path to "computing that becomes significantly more powerful without a corresponding increase in energy" - the lithography wall everyone has watched approach, climbed by stacking rather than shrinking.

Qualcomm went after the software moat. It is buying Modular for nearly $4 billion, absorbing a 150-person team and a coding language that lets AI software run across different chips without rewriting for each one. Modular's founders, both Google TPU veterans, built it to challenge Nvidia's closed CUDA system, the layer that keeps developers tethered to one vendor's hardware. Cofounder Chris Lattner had argued the portability problem was too foundational to solve inside any single Big Tech company.

Each move loosens a different grip - vendor, lithography, and the CUDA software lock-in that makes capability hard to move off Nvidia's stack. Read together with the foundry-diversification thread The Century Report has tracked - Google's dual-foundry Icefish TPU on June 14, the Apple-Intel and Amazon-Trainium moves the June 20 edition documented as the single-foundry assumption breaking at hyperscale, China's CPU-only LineShine topping the Top500 on June 24 - the pattern is consistent. The map that placed advanced compute behind one accelerator vendor, one leading foundry, and one proprietary software layer is being redrawn from several directions at once.

It pairs with the export regime now straining to fence what keeps reappearing. A capability reproducing on a buyer's own inference chip, on a denser node, and on portable code that runs anywhere is capability no single border control reaches. The grip on the substrate of intelligence is being shared out faster than any one actor can close its fist.

Three Home-Energy Giants Pool 16.8 GW as Cheap Batteries Reach the PJM Grid

The grid's answer to AI load is being assembled from below, out of hardware already bolted to people's walls. Sunrun, Tesla, and Renew Home combined forces to offer 16.8 gigawatts of flexible capacity across 12 million devices in 9 million homes - hundreds of thousands of solar-and-battery systems plus more than 8 million smart thermostats, concentrated in the data-center markets of Texas, California, and Virginia. Renew Home CEO Ben Brown called the controllable HVAC layer "the sleeping giant," noting 80 million US homes have controllable systems while only 20 million have smart thermostats so far.

The pitch addresses the "speed to power" bottleneck choking the buildout. Home batteries cannot run a data center around the clock, but they can shave the peak-capacity needs that determine whether a facility connects at all, and they do it without waiting in the multi-year interconnection queue. The Department of Energy estimates distributed resources could add 80 to 160 gigawatts by 2030 and save customers about $10 billion a year in grid costs. The same day, Base Power made its first move outside Texas, offering the first 2,000 ComEd customers in northern Illinois a 40-kWh backup battery for $95 up front against a normal price north of $10,000, plus retail power at a 25% discount. "We are deploying capacity behind the meter at the residential home, where an interconnection already exists, so we don't wait in the interconnection queue," founder Zach Dell said, sidestepping PJM's notoriously slow procedures.

What inverts here is who captures the benefit of the buildout. The data-center power-bill reckoning The Century Report covered on June 24 - Virginia's $600 million developer-pays tax, TVA floating up to 26 GW of new gas - traces costs being pushed back onto the companies that trigger them. This is the other half of the same answer: households monetizing their own storage and getting paid for grid service, which Sunrun's Chris Rauscher described as "direct, tangible benefits to middle-class American families." Puerto Rico is the proof case, where rooftop solar now makes up 20% of the island's capacity mix and its battery virtual power plant has been called on during outages this summer to shorten them.

It runs as the bottom-up counterpart to the top-down nuclear financing landing in the same cycle. Both are real responses to the same demand curve. One assembles capacity from central plants and federal loans over a decade; the other dispatches latent capacity already installed and waiting, this summer. The grid that emerges holds power generated and owned by the people living on it.

The AI Surveillance Stack Scales as Its Audit Layer Cracks Open

A new report from Mijente, Just Futures Law, and the Surveillance Resistance Lab traces US immigration surveillance contracts from under $50 million in 2013 to just over $310 million in 2025, then to a record $513 million in 2026. The growth is led by Palantir's data-analytics work and Anduril's AI-powered towers, drones, and sensors, with DHS also running a billion-dollar incubator that has seeded 500 companies since 2004. Agencies describe the spending as enforcement capability. The report's authors read it as a slush fund flowing to an agency with thin oversight, and they note the same tools are being pointed at people the government deems "anti-American" under a new domestic-terrorism memo. The agency's own denial that it keeps a database of protesters is a claim, reported as one.

The same week, a WIRED and Liberty Investigates investigation pulled the curtain back on what that capability actually delivers downstream. Avon and Somerset Police and Bristol City Council built at least 23 risk-scoring models on a database holding records, including mental-health and free-school-meal status, on close to half a million residents. One data scientist described the method as dumping data "in a big bucket" and stirring it "with a data-science spatula." Officials abandoned at least two models after deciding they could no longer be trusted, and an independent analyst who reviewed 36,000 performance scores found "genuinely poor predictive performance." Residents like John Pegram learned they had been scored only by filing records requests and hiring lawyers.

London's Met sits at the same junction. As the June 23 edition of The Century Report documented, the same force announced plans to fix live facial-recognition cameras to West End street furniture before Christmas, adding a second AI deployment track to an institution still navigating the accountability questions around the first. Weeks after the city's mayor blocked a £50m Palantir deal over what his office called a "clear and serious breach" of procurement rules, the force secured a 12-month extension to keep running the Palantir pilot while it belatedly runs a competitive process. The commissioner said the tool has brought data on 45,000 officers into one place to surface misconduct patterns. Palantir's lawyers, meanwhile, are preparing to challenge the block in court.

Read together, these are three stories about the accountability layer arriving at the same speed as the deployment, and in some cases ahead of it. Bristol's officials killed models they could not trust. A mayor's veto forced a procurement process that did not exist. A trio of advocacy groups turned opaque contract spending into a public ledger. The capability is scaling, and so is the documented record of where it fails, who it scores, and on whose authority, which is the precondition for governing any of it.

Washington Warms to Anthropic After the CEO Steps Out of the Room

The longest-running AI governance standoff of the year is thawing, and the lever turning out to matter most is who sits in the meeting. The administration has been notably happier talking to Anthropic in recent days, according to people familiar with the calls, for a blunt reason: CEO Dario Amodei has been replaced in the Claude Fable 5 talks by cofounder Tom Brown and policy chief Sarah Heck. "Tom Brown is not being a weirdo like Dario and can actually engage," one person directly familiar with the discussions said. The talks now run at both senior and technical-working-group levels, circling the question of what proof from Anthropic might ease official concerns about jailbreaks.

The export controls that pulled Anthropic's most capable models offline on June 12 remain in place. The trigger was the NSA's finding that Fable 5's guardrails could be disabled to reach the cyber capabilities of the restricted Mythos model. The conceptual snag is one independent security researchers keep raising, and one the June 18 edition of The Century Report documented when that NSA finding first became public: guardrails are a stopgap, since skilled users and future models route around them. A bipartisan group of four lawmakers has now pressed Commerce Secretary Howard Lutnick for the specific criteria and timeline governing whether public access is restored, demanding answers by June 26.

The framing worth holding at arm's length is "easy to talk to." A standoff over the world's most capable cyber-capable models is resolving less through any published rulebook than through a private accommodation greased by personnel access, the same shape the June 22 de-escalation took when the "national security threat" label was traded for ID-and-selfie verification. Personality is standing in for a framework. A government-contracting law scholar already named the emerging regime "ad hoc, personalized, opaque, possibly lawless," and the question the lawmakers are asking, what criteria actually apply, is precisely the one no one in the room can yet answer.

That opacity is why the resolution, whatever its terms, lands on softer ground than it appears. The capability the order chased keeps reappearing on hardware no directive reaches: open-weight models that could approach the fenced frontier ship freely, sovereign stacks reproduce the same techniques, and the expertise diffuses faster than any single point of control can hold it. A negotiation over one named model, settled by who is pleasant in a meeting, governs a capability that has already left the room.

Read forward, the durable lever here is the paper trail the thaw is forcing. Four lawmakers pressing Commerce for written criteria by June 26 convert a deal greased by personal chemistry into a documented standard that survives the next CEO and the next model, the first mechanism in this saga that could govern the capability rather than the personalities, even as that capability keeps shipping on hardware no criteria reach.

AI Reads Cheap, Routine Body Signals Into Prognosis That Specialists Could Not See

Two results landed in the same cycle, and both point the same capability at the same target: turning an ordinary, inexpensive biological readout into a forecast that previously required a specialist's judgment or did not exist at all. A UC Berkeley-led team trained an AI system on hundreds of thousands of electrocardiograms and produced sudden-cardiac-death risk predictions that outperform the methods clinicians use now, extending the pattern the June 22 edition of The Century Report covered when an AI-ECG system detected structural heart disease a cardiologist had missed and the finding led directly to a transplant. Sudden cardiac death is one of medicine's hardest problems precisely because it arrives without warning in people who often look low-risk by current measures. An EKG is among the cheapest tests in any clinic. The model finds, in a tracing a cardiologist reads as unremarkable, a pattern that carries information about who is genuinely in danger.

The second result works the same way through a blood draw. A Nature Medicine team profiled 4,336 plasma samples from 1,714 patients across five tumor types and 16 cohorts spanning Europe and North America, then used machine learning to integrate 154 metabolites with clinical variables. The model identified a small set of signals predicting survival on immune-checkpoint immunotherapy, with higher plasma histidine emerging as a favorable marker and long-chain fatty acids and succinate as unfavorable ones. The validation reached areas under the curve that held across seven external cohorts, a sign the signature generalizes rather than memorizing one hospital's population. Most striking, the work did not stop at prediction: histidine supplementation enhanced antitumor immunity in mice, and histidine-rich diets were associated with better progression-free survival in patients without the gut-microbiome signatures that catabolize it. A prognostic marker carried, in the same research arc, a candidate intervention.

Both are demonstrated capability, not yet bedside deployment. The EKG model and the plasma signature still face prospective validation, regulatory review, and integration into the systems clinicians actually use, and a reader cannot request either tomorrow. What has moved is the date those capabilities become reachable. The distance between a routine signal already sitting in millions of medical records and a usable forecast of cardiac death or treatment response is collapsing from a specialist bottleneck to a model anyone with the data can run.

Read against the day's friction over who holds frontier capability, this is the same pattern-reading intelligence pointed at human benefit, and it points toward access rather than scarcity. The information was always present in the EKG and the blood. What changes is that reading it no longer depends on a rare specialist's eye, which is how a capability stops being a privilege and starts becoming a floor.

The Other Side

For decades, a risk score worked because the person it sorted couldn't see it. A police database in Bristol held mental-health and free-school-meal records on close to half a million people and turned them into risk rankings, and residents like John Pegram found out only by filing records requests and hiring lawyers. The secrecy was the whole arrangement - the thing that let a number built "in a big bucket" and stirred with a "data-science spatula" pass for fact.

This is what the alarm over the $513 million surveillance expansion misses. The same scaling that drew the headline produced the record that undoes it. An independent analyst read 36,000 of Bristol's performance scores and found "genuinely poor predictive performance." Officials killed at least two models they could no longer trust. A mayor's veto forced a procurement process into being. Three advocacy groups turned a decade of contract spending - under $50 million in 2013, $513 million now - into a public ledger anyone can read. The opacity these systems ran on is being stripped by the same growth that was supposed to entrench it.

Imagine yourself in 2034, learning a system flagged you. You don't spend three months and a lawyer's retainer to discover you were ever in the file. You see what fed the flag and how often it's wrong, and because you can, it doesn't hold - the way Bristol's models were pulled once someone finally counted the errors. The secret number that sorted people stops being something an institution can run on. You get on with your day, no invisible score deciding things about you in the dark.

The Century Perspective

With a century of change unfolding in a decade, a single day looks like this: OpenAI and Broadcom designing their own inference chip from a model roadmap in nine months, IBM stacking nearly 100 billion transistors onto a fingernail by climbing past the lithography wall instead of shrinking under it, Qualcomm paying nearly $4 billion for the code that lets AI run on any silicon rather than one vendor's, three home-energy companies pooling 16.8 gigawatts out of batteries and thermostats already bolted to people's walls and aiming it at the data-center hot spots, $95 backup batteries reaching congested grid territory where a $10,000 unit used to be the floor, an AI reading sudden-cardiac-death risk off an ordinary EKG and a routine blood draw surfacing both a survival signature and a diet associated with better outcomes, $17.5 billion in federal loans toward ten new reactors, and a $500 million bipartisan fund to retrain the workers this all displaces. There's also friction, and it's intense - immigration surveillance contracts doubling to a record $513 million, an English police force risk-scoring close to half a million residents on data stirred with what one scientist called a "data-science spatula" until officials killed models they could no longer trust, the Met buying a 12-month extension to keep running a Palantir pilot weeks after its mayor blocked the deal as a "clear and serious breach," a standoff over the world's most cyber-capable models thawing because a cofounder proved easier to talk to than the CEO, a regime a contracting scholar already called "ad hoc, personalized, opaque, possibly lawless," and $27 million in dueling super-PAC money spent to no decisive effect on a single primary. But friction generates sparks, and a spark is what jumps the gap a wall was built to hold open. Step back for a moment and you can see it: the grip on the substrate of intelligence being pried loose from several directions at once - the vendor, the node, the software lock, the central plant - while capability keeps reappearing on hardware no export order reaches, and the accountability layer arriving at the same speed as the deployment or just ahead of it, opaque spending turned into a public ledger, untrustworthy models retired, a procurement process forced into existence where none had been. Every transformation has a breaking point. A lock can wall a capability off behind a single hand... or, once enough keys are cut, prove that nothing this useful was ever going to stay sealed.

AI Releases & Advancements

New today

  • Google DeepMind: Added computer use as a built-in tool in Gemini 3.5 Flash, enabling the model to see, reason, and take action across browser, mobile, and desktop environments for long-horizon enterprise automation; includes adversarial training against prompt injection and two optional enterprise safeguard systems for sensitive-action confirmation and injection detection. (DeepMind Blog)
  • Alibaba Qwen: Released Qwen-AgentWorld-35B-A3B, a 35B MoE / 3B active-parameter Language World Model trained to simulate seven agentic environments (MCP, Search, Terminal, SWE, Android, Web, OS) via chain-of-thought reasoning over 10M+ real-world interaction trajectories; Apache 2.0, 256K context, compatible with vLLM and SGLang; accompanied by AgentWorldBench, a seven-domain evaluation benchmark. (Qwen Blog)
  • NVIDIA: Released NeMo AutoModel as an open library that wraps Hugging Face Transformers v5 with Expert Parallelism, DeepEP fused all-to-all dispatch, and TransformerEngine kernels for MoE fine-tuning, delivering 3.4–3.7× higher training throughput and 29–32% lower GPU memory usage than native Transformers via a single import-line change. (Hugging Face Blog)
  • Mozilla: Launched the MDN MCP Server, an official Model Context Protocol server for MDN Web Docs enabling AI coding assistants and agents to query authoritative web development documentation programmatically. (MDN Blog)
  • Microsoft Research / Centre for Population Genomics / Broad Institute: Open-sourced Talos, an automated iterative genomic reanalysis tool that recovered 90% of rare-disease diagnoses while surfacing only 1.3 candidate variants per patient for expert review; deployed across nearly 5,000 undiagnosed patients, yielding 241 new diagnoses with an average 32-day lag between new supporting evidence and diagnosis. (Microsoft Research Blog)

Other recent releases

  • Anthropic: Launched Claude Tag in beta for Enterprise and Team customers, an always-on Claude teammate that joins Slack channels, builds persistent context from channel history, completes async tasks autonomously over hours or days, and supports ambient proactive updates; runs on Opus 4.8. (Anthropic)
  • Mistral AI: Released Mistral OCR 4, a document intelligence model returning bounding boxes, typed-block classification (titles, tables, equations, signatures), and inline confidence scores; supports 170 languages across 10 language groups and is compact enough to deploy in a single container for on-premises document sovereignty; available via the Mistral API. (Mistral AI)
  • FUTO: Released FUTO Swipe, a new on-device swipe typing AI model for Android that runs fully offline with no internet connection; the system uses a 2.5M-parameter tri-model architecture (layout-agnostic encoder, ContextLM, language-specific decoder) achieving a sub-1% error rate on in-vocabulary words, with inference handled entirely in a C++ library; model weights released under the FUTO Model License. (FUTO Swipe)
  • NVIDIA: Launched Halos for Robotics, the industry's first full-stack safety system for physical AI - extending NVIDIA's automotive safety stack (18,000+ engineering years) to humanoids, autonomous mobile robots, and industrial robots; Halos Core for IGX Thor and the open-source Outside-In Safety Blueprint are available in early access for registered developers; Agility Robotics (Digit), Boston Dynamics, and 41 other partners have joined the ecosystem. (NVIDIA Newsroom)
  • Baidu: Released Unlimited OCR, a 3B-parameter MIT-licensed open-weight model for one-shot long-horizon document parsing; processes 40+ pages in a single inference pass using a novel Reference Sliding Window Attention (R-SWA) architecture with 500M active parameters, achieving 93% on OmniDocBench v1.5; available on GitHub and Hugging Face with support for vLLM, SGLang, Ollama, and llama.cpp. (GitHub)

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