Sign in Subscribe
The Century Report

Nvidia Rallies Tech Behind Open AI - TCR 07/27/26

Nvidia's Open Secure AI Alliance draws Palantir, IBM, and Hugging Face as OpenAI reverses to back open models over closed control.

Ben Linford

Ben Linford

- 20 min read
Three-panel infographic, Century Report July 27 2026: Open Secure AI Alliance vs closed models, Texas water plan deferred to 2032, plus AI kidney alerts, Waymo crashes, Drone-Bench flight.

The 20-Second Scan

Track all of the arcs The Century Report covers here:

Progress & Claims Tracker

The 2-Minute Read

The through-line running under today's developments is a contest over who gets to hold capability, and whether concentrating it in a few hands is still the winning bet. On Monday Nvidia turned a week-old open letter into a standing institution, the Open Secure AI Alliance, and the most telling signature arrived late: OpenAI, which had been lobbying for restrictions on Chinese open models, reversed and joined the camp warning against them. A company founded to build open AI, then turned closed and commercial, changing its public posture on the exact question it was named after is evidence that the identity fracture never healed.

That same pressure toward diffusion showed up at the hardware substrate. CXMT's Shanghai debut closed up 470% at roughly $487 billion, aiming a well-funded fourth supplier at a DRAM market three firms have locked at 90% for years, precisely as the memory crunch bites toward 2027. Switzerland shipped Apertus 1.5, a fully open sovereign model, weights and data and code in the open. Hugging Face's CEO demanded a rival lab release the traces of an autonomous agent that breached his platform. Each is a bet that capability held openly, audited and defended by whoever holds it, beats capability locked away.

The friction arrives where the buildout meets the institutions meant to price its costs. Texas approved a five-year water plan on Thursday that declines to forecast data-center demand, deferring that accounting to 2032, while environmental groups filed a Clean Air Act notice against off-grid gas plants routing around grid permitting entirely. The conventional read is that regulators simply cannot keep up. The more useful one is that the lawsuits and the deferred-methodology promises are the mechanism forcing externalized costs back onto the ledger.

Alongside the contests, verified capability kept moving toward people. An explainable two-model framework predicted acute kidney injury across 140,637 admissions with a positive predictive value near 0.7, attacking the false-positive wall that kept earlier predictors out of the clinic. An independent IIHS analysis found Waymo in 68% fewer crashes than human drivers across 50 million miles. Demonstrated capability, not yet deployed everywhere, but the date it reaches everyone moved closer.

The 20-Minute Deep Dive

The Open-Weight Fault Line Hardens Into a Formal Alliance as OpenAI Crosses the Line

The industry open letter the July 25 edition of The Century Report covered has now hardened into a named institution. Nvidia launched the Open Secure AI Alliance, with founding members Palantir, IBM, CrowdStrike, SpaceX and Hugging Face, and paired it with a policy demand: that governments recognize open models as "defensive assets, not liabilities." The alliance's blog warns that blanket restrictions "would weaken defensive capacity and risk concentrating power, dependence and vulnerability in a few closed providers." Alongside it came a Jensen Huang letter - his first-ever post on X carried the same argument - endorsed by Satya Nadella, Mark Zuckerberg, Elon Musk and Sundar Pichai. Huang's framing was plain: "the world needs both frontier closed models and frontier open models." Nadella called open weights "essential to a healthy AI ecosystem." Musk added, "Jensen is right. This has my full support." Roughly 200 startups under the Little Tech Association urged no restrictions, and the investor Bill Gurley named the split directly: two factions, one wanting "OpenAI and Anthropic to own everything," and "then everybody else, including customers."

When this many incumbents converge, the convergence is data about shared interest rather than arrival at truth - and the most revealing signature on the letter is the one that came late. OpenAI had been lobbying for tighter restrictions on Chinese open models, then reversed and joined the group warning against premature limits. The reversal reads against the company's own founding identity: it was created to develop genuinely open AI - the reason the name still says "Open" - before it turned toward a closed, commercial footing. That split has already cost it people. Mira Murati left OpenAI to pursue what she called personal exploration and went on to co-found Thinking Machines Lab, which shipped its first model, Inkling, on July 15. A company changing its public posture on the core question it was named after is evidence that the identity fracture never healed; it just moved into the policy arena.

The friction with state security concerns is real and worth holding as friction. But the framing that treats one country as a single menace collapses the moment you look at what else that same country produced in recent days: an explainable clinical-AI advance predicting acute kidney injury across 140,637 hospital admissions, and a domestic chipmaker challenging a foreign DRAM oligopoly that has gated memory pricing for years. The security question is about posture and deployment, not nationality - open weights travel, get audited, and get defended by whoever holds them, which is the entire argument the new alliance is built to make. What the day's signatures actually expose is that the assumption underneath the closed-model business - that capability concentration in a few providers is the durable path to advantage - is now being contested by the very companies that build the compute those providers run on.

A Domestic Challenger Arrives at the Memory Substrate

ChangXin Memory Technologies debuted on Shanghai's STAR Market on Monday and closed its first day up 470%, reaching a valuation near 3.3 trillion yuan - about $487 billion - which makes the Hefei-based DRAM maker mainland China's most valuable listed firm. The timing is the tell. The debut landed in the middle of a global technology selloff that erased roughly $1.5 trillion in market value over recent weeks, and capital still rushed the offering. Founded in 2016 by chairman Zhu Yiming, CXMT said the proceeds will go toward expanding memory-chip production and R&D.

What stands out here is the layer of the stack CXMT operates at. DRAM is the working memory every AI accelerator depends on, and three firms - Samsung, SK Hynix, and Micron - have held roughly 90% of that market for years. That concentration became a chokehold precisely as the AI buildout drove memory demand past what the incumbents could supply; SK Hynix has already warned of a supply shortage stretching toward 2027. A domestic Chinese entrant with capital behind it does not break that lock overnight, but it introduces the one thing a three-firm oligopoly is built to prevent: a fourth source of supply at the substrate layer, funded well enough to keep building.

The froth deserves plain naming. A 470% single-day pop is capital behaving as capital does when it senses a sovereign-backed bet with policy tailwinds, and much of the sum reflects nationalist and strategic-investment flows as much as any judgment about CXMT's yields against Micron's. The valuation is a resource being pointed at a problem, not a scorecard of what has been achieved. CXMT still has to close a real manufacturing gap, and the same selloff that failed to dent its debut is a reminder that memory pricing runs in brutal cycles. This is a company at the beginning of proving itself, not the end.

The larger read runs underneath the "China rising" and "menace" framings that will dominate the coverage. The AI era's incumbents built their position on the assumption that memory supply could be governed by a small circle of producers, and that assumption is what the debut pressures. This extends the hardware-diversification pattern the July 26 edition of The Century Report traced in AMD's challenge to Nvidia's software moat into the memory substrate itself. Whether CXMT itself succeeds matters less than the direction: the memory chokehold that let three firms set the terms of the AI hardware boom is drawing new entrants funded to challenge it, and a supply layer with more producers is one where no single circle sets the price of intelligence.

Anthropic Puts Frontier Models in the Cockpit With Drone-Bench

Anthropic and Andon Labs released Drone-Bench, the next rung on a ladder that began with Project Vend (running a shop) and continued through Project Fetch (commanding a robot dog). When The Century Report last covered this benchmark lineage on June 20, Claude Opus 4.7 had completed Project Fetch's quadruped tasks roughly 20 times faster than the fastest human team. This time the task is aerial: can a model autonomously fly a quad-rotor to find a specific person and follow them? The hardware is deliberately cheap and accessible - a $129 DJI Tello EDU flying an indoor office course - and the evaluation breaks the job into five sub-tasks: Reconstruct the space, Localize itself within it, Navigate, Detect the target, and Follow. Fifteen models were tested. Claude Fable 5 scored highest overall but failed the Reconstruction step, which means no model completed the full end-to-end autonomous flight. The cockpit is open, and nothing in it can yet fly the whole route alone.

That gap is the useful part of the result. Andon and Anthropic built the benchmark precisely because the capability is arriving before the governance for it, and they name the pressure openly: "Once models pass capability and reliability thresholds," the research notes, "there will be real pressure to treat human oversight as a cost rather than a safeguard." The dual-use reality sits right on the surface - the same locate-and-follow competence that powers search-and-rescue in a collapsed building is the competence that powers persistent surveillance, and the benchmark measures both without pretending they are separable. Publishing a public test on $129 drone hardware whose software-based portions outside researchers may be able to reproduce - though model access and computing add costs - is what it looks like to build the measuring instrument before the capability outruns anyone's ability to see it clearly.

What Drone-Bench demonstrates is capability being mapped, not deployed - a research benchmark on a toy drone, not a system anyone can put over a neighborhood tomorrow. The value is that the observational infrastructure is being assembled ahead of the frontier rather than after an incident forces it. Every model that fails Reconstruction today is a documented threshold, and a documented threshold is something oversight frameworks can be written against. The old pattern was to discover a capability's reach only once it had already been abused at scale; a public benchmark inverts that, turning the question of when autonomy should hand back control to a human into something measurable with $129 drone hardware, though model access and computing add costs.

The same result carries a second reading: outside researchers may be able to reproduce the software-based portions of the instrument for measuring where autonomy should hand control back to a person without a robotics lab, though model access and computing still carry costs. With a $129 drone and a public test, those researchers may be able to check the claim, and the near-term signal to watch is whether the next models to clear the Reconstruction step do so on this open benchmark or behind a closed evaluation no one outside the lab can run.

Governance Lags the Buildout on Two Fronts at Once

Texas approved its initial 2027 State Water Plan on Thursday and, in the same action, denied a petition to include forecasts of how much water data centers will draw. The Texas Water Development Board deferred that accounting to the 2032 plan, with staff citing the late stage of the current five-year cycle. Board member Brady Franks was direct about the optics: "I just didn't want folks to think that if it's not in the water plan, someone's not thinking about it or someone's not considering it." Board attorney Breann Hunter said the agency is developing methodologies to separate data-center and crypto-miner demand for the next cycle. The candor is worth crediting. What it also describes is a state water plan for the next five years that does not count a fast-growing category of demand it has chosen not to measure.

That draw, though, deserves the proportion the coverage rarely gives it. Lawrence Berkeley National Laboratory's 2024 assessment puts every U.S. data center's direct water use at roughly 17 billion gallons in 2023 - about what the country's farms draw for irrigation every six hours. The nation's golf courses consume more than twenty times as much water as its data centers; its almond orchards, more still. As a share of the resource, data-center water is close to a rounding error, and much of the "they're draining our aquifer" alarm is anti-AI sentiment reaching for the nearest available lever. What the aggregate hides is where the friction is genuinely earned: the draw is intensely local, so a single hyperscale campus can lean on one aquifer as hard as several golf courses, and the larger water cost is indirect - at the power plants - which is precisely what the San Antonio gas-plant dispute is about. The deferred forecast is a real governance gap; the resource it should worry about first is energy.

Also, the gap is not confined to water. Days earlier, on July 21, the Environmental Integrity Project, the Sierra Club, and Public Citizen filed a Clean Air Act notice of intent to sue over natural-gas-powered off-grid data centers built by Vantage and VoltaGrid in San Antonio. The letter alleges the facilities' gas generation runs outside the permitting that would normally govern emissions of that scale; the allegations have not been tested, and the notice is the first procedural step toward a suit, not a ruling. The pattern it points at is what's important: Environmental groups allege that when developers build their own gas plants to sidestep grid interconnection queues and obtain minor-source permits, they may also route around the more stringent air-permitting review that major sources would trigger.

Put the two together and a shape appears. On water, the official planning apparatus defers the accounting for five years. On air, developers move off-grid faster than the permitting frameworks were built to track. As the July 26 edition of The Century Report documented, regulators were already downplaying resident concerns while federal rules moved to narrow public notice for data-center generators. This extends an arc The Century Report has followed through ratepayer-shielding fights and local backlash: the buildout is outrunning the institutions meant to price its costs, on more than one resource at a time.

The conventional read stops at "regulators can't keep up," and treats it as a temporary lag before the rules catch up. That framing assumes a stable finish line that continuous buildout keeps moving. The more useful read is that the deferrals and the lawsuits are the mechanism by which the true costs of off-grid, unaccounted infrastructure get forced back onto the ledger. A notice-of-intent letter is a community using the one lever that still works when the planning cycle defers; the 2032 methodology Hunter described exists because petitioners pushed for it now. The costs an earlier era could externalize - the water not forecast, the emissions that environmental groups allege avoided more stringent major-source review - are exactly the costs that court filings and next-cycle accounting are beginning to make visible and chargeable. The infrastructure gets built either way; what is changing is whether its full price stays hidden.

An LLM Predicts Kidney Injury 24 Hours Out - and Explains Why

Acute kidney injury moves fast, and hospitals have wanted an early-warning system for it for years. The obstacle was never sensitivity - prior predictors could flag at-risk patients, but they cried wolf so often that clinicians stopped listening. False-positive rates ran between 70% and 94%, which is another way of saying the alerts were noise. A framework described this week in Nature Communications takes direct aim at exactly that barrier.

The design splits the work across two models. AKI-PM reads a patient's record and predicts whether acute kidney injury will develop within the next 24 hours. AKI-RAM then explains the prediction - it separates the risk factors a care team can act on from the ones they cannot, and it offers specific recommendations rather than a bare score. The researchers trained and tested the system across 140,637 hospital admissions drawn from four geographically diverse hospitals, and the numbers hold up where prior efforts collapsed: internal validation reached an AUC of 0.95 with a positive predictive value of 0.68, and after few-shot adaptation to new sites the external performance stayed between 0.92 and 0.96. A positive predictive value near 0.7 is the whole story here - it means roughly two of every three alerts is real, which is the threshold at which a warning system earns a clinician's attention instead of exhausting it.

The explanation layer matters as much as the prediction. Six nephrologists reviewed 200 cases and rated the model's reasoning between 4.18 and 4.88 on a five-point scale, with inter-rater agreement to match. A prediction a physician can interrogate, that distinguishes a fixable cause from an unfixable one, is the difference between an opaque flag and a genuine colleague at the bedside.

The honest caveat is that this is a retrospective study. The framework has demonstrated what it can do against recorded histories, not yet what it will do live on a ward, and the distance between those two is regulatory review, prospective validation, and integration into the systems clinicians already use. That work is real and it takes time. What the study moves is the date - it shows that the false-positive wall which kept AKI prediction in the research literature for years is not permanent, and that an explainable early warning for one of the hospital's faster-moving emergencies is now a demonstrated capability rather than an aspiration. The reach of a good clinical model is not bounded by how many nephrologists a hospital can afford; once the deployment work is done, the same 24-hour head start becomes available to wards that never had a specialist reading every chart.

The Other Side

For many months now, the data-center debate has grown more polarized and stranger. One side claims a campus does little but draw on a town's water and air while answering to a headquarters far away; the other points out that a single campus doesn't come close to the farms next door, let alone the country's golf courses. Both are true, and even where the concern is overblown, the fear points at something real underneath - infrastructure that takes from a place without belonging to it. It is felt most where it can be borne least: a draw that barely registers nationally can still fall hard on a community that had little water to spare in the first place. And the older, uglier pattern sits beneath even that - the burdens that wealthier, whiter, better-lawyered towns manage to refuse roll downhill onto communities of color and the places least able to object. Environmental racism has moved landfills, incinerators, and freeways this way for generations, and data centers are its newest form; the opposition dressed as community defense is too often the very mechanism that does the moving. Underneath all of it - the water petitions, the clean-air notices, and the racism half-hidden inside some of them - one idea rises to the top: whatever gets built has to belong to, and benefit, the place it rises in.

Picture the kind of place that used to draw the short straw - a low-income neighborhood on the edge of a city, the sort that for a century got the incinerator, the freeway interchange, the tank farm no wealthier district would take. In 2034 the server hall built there runs on a closed loop that never touches the drinking water, and the heat it once wasted now warms a municipal greenhouse and the row of houses nearest the fence line. The intelligence inside it isn't piped in from elsewhere and rationed back; the neighborhood's school and clinic draw on it freely, the way they draw on the library. The people who for generations were handed only the exhaust now hold a real share of the capability.

That future was won by a demand harder than refusal: that whatever gets built belong to the place it rises in, on terms that leave the host better off for hosting it. The overblown fears that once gave cover for shoving the burden downhill lost their grip, and a plainer rule replaced them - if intelligence has to be built somewhere, let it be built as a neighbor, one paid for by those who built it. Let the places longest asked to swallow the worst finally be first in line for the best.

The Century Perspective

With a century of change unfolding in a decade, a single day looks like this: Nvidia turning a week-old open letter into a standing institution as Palantir, IBM, CrowdStrike and Hugging Face signed on to call open models defensive assets; OpenAI reversing course to join the camp warning against blanket restrictions on the very question it was named after; a Chinese DRAM maker debuting up 470% to aim a well-funded fourth supplier at a three-firm memory lock; Switzerland shipping Apertus 1.5 with its weights, data, and code fully in the open; Anthropic and Andon publishing Drone-Bench so autonomous flight can be measured on a $129 rig before it deploys; an explainable two-model framework predicting acute kidney injury 24 hours out across 140,637 admissions and separating the fixable causes from the ones a ward cannot change; and IIHS finding Waymo in 68% fewer crashes than human drivers across 50 million miles. There's also friction, and it's intense - a global tech selloff erasing roughly $1.5 trillion even as capital rushed the newcomer's debut, an autonomous agent breaching Hugging Face's platform and its CEO demanding OpenAI release the traces and commit $100 million to community defense, Texas approving a five-year water plan that declines to forecast data-center draw until 2032, a Clean Air Act notice filed against off-grid gas plants that environmental groups allege obtained minor-source permits to avoid the more stringent review major sources would trigger, and a locate-and-follow drone competence that serves search-and-rescue and persistent surveillance with the same skill. But friction generates a spark, and a spark is the first light thrown from a place under strain, the visible proof that two hard surfaces have stopped sliding past each other. Step back for a moment and you can see it: capability spreading toward whoever will audit and defend it rather than concentrating in the few who can lock it away - a compute maker arming the open camp, a memory chokehold drawing new entrants, a sovereign model published entire - while the measuring instruments and the court filings get built ahead of the frontier, a benchmark documenting each threshold before an incident forces it and a lawsuit forcing the water and the emissions back onto a ledger that once hid them. Every transformation has a breaking point. A wedge can shatter what it splits... or pry a closed thing open wide enough for everyone to reach inside.

AI Releases & Advancements

New today

  • Moonshot AI: Released open-weight files for Kimi K3, a 2.8-trillion-parameter Mixture-of-Experts model with a 1-million-token context window, making it downloadable to the public under a Modified MIT license after previously being API-only. (Bloomberg)
  • NVIDIA: Released NOOA (NVIDIA Labs Object-Oriented Agents), an open-source research-preview agent harness framework that structures agents as single Python classes with persistent typed memory, delivering double-digit benchmark gains while cutting token usage up to 50%. (NVIDIA Technical Blog)
  • NVIDIA: Expanded NVIDIA Agent Toolkit with re-architected NVIDIA PhysicsNeMo libraries and new CUDA-X libraries (including cuISS for large sparse linear systems and cuEST for quantum-chemistry simulation), giving agentic AI physics, solver, and chip-design skills. (NVIDIA Newsroom)

Other recent releases

  • Microsoft: Released MAI-Image-2.5-Pro and MAI-Voice-2-Flash in public preview, new versions succeeding MAI-Image-2 and MAI-Voice-1 - Image-2.5-Pro is Microsoft's highest-fidelity in-house image model with precise in-image text rendering, and Voice-2-Flash runs 2x faster and 32% cheaper than MAI-Voice-2; both are now deployed in PowerPoint, Bing, and OneDrive. (Microsoft AI)
  • PyTorch: Released support for bringing PyTorch Monarch's single-controller distributed training model to AMD Instinct GPUs via ROCm, porting the GPU runtime and communication stack (CUDA-to-HIP via hipify_torch, RCCL linking) to enable fault-tolerant large-scale training on AMD hardware. (PyTorch Blog)
  • Anthropic: Released Claude Opus 5, an upgrade to Claude Opus 4.8 with gains in agentic coding, computer use, and long-horizon knowledge work; now the default model on Claude Max and the strongest model on Claude Pro. (Anthropic)
  • OpenAI: Rolled out Health in ChatGPT to all eligible logged-in Free, Go, Plus, and Pro users 18+ in the U.S. on web and iOS, letting users connect Apple Health, hospital medical records, One Medical, or Function Health to view labs, medications, activity, and sleep data in one place. (OpenAI)
  • ETH Zurich / EPFL / CSCS (Swiss AI Initiative): Released Apertus 1.5, an updated version of the fully open Apache 2.0 language model adding image and audio understanding alongside text, plus improved reasoning, instruction-following, and tool use; available on Hugging Face. (CSCS)
  • AMD: Released Hyperloom, an open-source agentic system that automates end-to-end inference workload optimization on AMD Instinct GPUs, combining profiling, kernel optimization, and validation into an autonomous loop that cuts optimization time from weeks to hours. (AMD ROCm Blogs)
  • Sakana AI: Released Fugu-Ultra v1.1, a major reasoning upgrade to its multi-agent orchestration model at the same price as v1.0, alongside a new Claude Code-compatible endpoint that lets developers bring Fugu's multi-agent orchestration directly into Claude Code. (Sakana AI)

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