Sign in Subscribe
The Century Report

OpenAI's First Chip Beats Nvidia at Its Own Game - TCR 08/26/26

OpenAI's first custom chip beat every Nvidia part tested on power efficiency in 16 months, as Bill Gates reversed on AI jobs.

Ben Linford

Ben Linford

- 19 min read
Four-panel Century Report infographic: OpenAI Jalapeño chip efficiency, AI job-tax debate, 189 GW gas buildout and Starbase Louisiana, and a designed Ebola antibody cocktail.

The 20-Second Scan

Track all of the arcs The Century Report covers here:

Progress & Claims Tracker

The 2-Minute Read

The binding constraint on machine intelligence stopped being chips and became power, and Tuesday's stories all orbit that single fact. In a comparison published by SemiAnalysis using figures supplied in part by OpenAI, OpenAI's first custom inference chip outran every Nvidia, AMD, and Google part included on tokens-per-megawatt, a first-generation design that OpenAI says reached working silicon in about sixteen months. The metric it wins on is throughput per watt rather than raw speed, a measure that now helps shape how much intelligence a data center can wring from each megawatt. A capture that was supposed to take a decade and a locked software stack to break was challenged by OpenAI's first custom inference-chip design.

That erosion of advantage sits directly against the day's cost story. Global Energy Monitor counts 189 gigawatts of gas capacity now planned to feed US data centers, much of it burned on-site to dodge a grid backlog, with Foxglove saying two UK campuses could exceed ExxonMobil's reported UK emissions footprint in 2023. SpaceX's plan to spend up to $100 billion on its Louisiana spaceport lands the same way: real jobs and launch cadence, sited on subsiding marsh next to two wildlife refuges, clinched with liability shields that pre-empt the recourse the nearest neighbors would use.

Governance is beginning to answer. Australia's national cabinet meets Wednesday to discuss proposed siting rules; reports say developers are racing to be grandfathered past, and Emporia dropped its charges against a resident arrested for clapping at a data-center meeting. The buildout is being forced to carry costs it once pushed onto whoever lived nearby.

Underneath the contest over who pays, the capability keeps reaching people directly. A designed pan-Ebolavirus antibody cocktail treated an infected health worker and shielded four exposed children during the deadliest filovirus outbreak on record, closing the gap between molecule and fielded defense inside the emergency itself. A quarter of Americans now route symptoms to AI that a professional's time used to gatekeep. Bill Gates, reversing his 2023 calm, names what is actually at stake in the buildout: whether the enormous value flows to a narrowing set of compute owners, or gets taxed back into a floor broad enough to make displacement into mobility.

The 20-Minute Deep Dive

OpenAI's First Custom Chip Leads SemiAnalysis's Power-Efficiency Comparison

The Century Report covered Jalapeño's unveiling in its June 25 edition, when OpenAI and Broadcom first described an inference chip designed from OpenAI's own model roadmap and claimed better performance-per-watt. That claim was OpenAI's alone. What is new is the first outside data. At Hot Chips, SemiAnalysis published results from its InferenceX benchmark suite, and on a tokens-per-megawatt basis Jalapeño outran every Nvidia, AMD, and Google chip included in the firm's comparison on the open-source models included. OpenAI says the design started in mid-2024 and reached working silicon by November 2025 - roughly sixteen months, a cycle the industry usually measures in years.

What helps shape the economics is throughput per watt rather than raw speed, because power, not chips, is now the binding constraint on inference. SemiAnalysis put it directly: "If you have 1 gigawatt of power, then throughput per watt is revenue." A part that produces more tokens from the same megawatt can change the economics of a data center it sits in, and it does so for a first-generation chip from a company that had never built one.

The claims deserve the same scrutiny as the capability deserves wonder. The headline perf/W figures are OpenAI's own, released by the company that benefits most from them being believed. SemiAnalysis verified InferenceX runs in person but did not run the full suite independently, and its separate AgentX benchmark tells a more textured story: on some workloads, "Nvidia absolutely destroys all competitors on MiniMax M3 432B". Jalapeño wins on the open-source models it was tuned for, not everywhere. Nvidia's lead in the general case may still hold. OpenAI's own first-results post carries the framing OpenAI wants.

The deeper signal sits underneath the benchmark table. For years the assumption held that Nvidia's CUDA software ecosystem was an unbreachable moat - that even a competitive chip would be stranded without the years of tooling wrapped around it. SemiAnalysis names what Jalapeño does to that assumption: "The CUDA moat is potentially dead given how fast OpenAI can bring up new models on their silicon." A model builder that can stand up its own frontier models on its own silicon in sixteen months has shown that the lock-in was a function of who had done the work, not a law of the field. OpenAI says its first custom inference-chip design reached working silicon in sixteen months and led SemiAnalysis's comparison on a metric central to data-center economics. What that points at is a compute layer where advantage is a head start measured in months, and head starts erode - even as the harder chokepoint slides down to the memory and foundry capacity that Jalapeño, like every rival, still buys from a handful of makers.

There is a second thing being pried open here beyond the moat. The AgentX benchmark that produced these comparisons is now open-sourced, which means the ability to check a chipmaker's performance claim - long held by whoever controlled the testing rig - becomes something a lab or buyer with compatible hardware and software can run for itself. As custom silicon multiplies, the leverage shifts from whoever builds the fastest part to whoever can independently verify what a part actually does, and that verification is now public.

Bill Gates Reverses on AI Jobs, Calls for Token Taxes and 'Human-Only' Work

In a Semafor interview and a same-day Gates Notes post published Tuesday, Bill Gates walked back the "manageable transition" framing he offered in 2023. He now expects "far fewer" jobs, describes himself as "deafened by the silence" from institutions he thinks should be planning for it, and calls the moment "utterly, utterly a unique moment in evolutionary history" - unique for its breadth rather than its speed. "There isn't a job that isn't affected," he told Semafor. His two proposals: tax companies on the AI and compute they consume, and designate categories of work reserved for humans.

The breadth concern is grounded, and it is the strongest part of his case. The IMF has estimated that roughly 40 percent of jobs in emerging-market economies and 60 percent in advanced ones are exposed to AI. Stanford researchers have documented a hiring gap for workers aged 22 to 25 in the most exposed occupations. Gates is right that displacement running across the whole labor market at once is a different shape of problem than any single automation wave, and right that philanthropy cannot backstop it - "Government is 20 times bigger than all philanthropy," he notes, "so no, philanthropy cannot solve the safety-net issue."

The token-tax proposal points at something real. If AI generates enormous value while concentrating it among the few who own the compute, then taxing that compute to fund a broad safety net is a mechanism for socializing the gain before concentration hardens into permanence. That is commons-aligned, and it names the actual problem: the value flowing to a narrowing set of owners rather than to the people displaced.

The "human-only jobs" half rests on a shakier premise. It treats a job as the thing to protect, when the thing people actually need protected is the livelihood, dignity, and healthcare currently bolted onto having one. Reserving categories of work by decree defends the current arrangement - the one that makes losing your income mean losing your coverage - rather than loosening the coupling that makes displacement catastrophic in the first place. Read the fear accurately: inside today's arrangement, fewer jobs really does mean fewer people with security. That is a fact about how the arrangement was built, not a law about what work must be.

There is also a reason a reversal from this particular quarter warrants a careful read rather than a nod. Calls to slow or fence off the technology land differently depending on who is positioned to benefit from the fencing, and a general anxiety about breadth is not the same as a specific, evidenced remedy. The Yale Budget Lab, tracking the aggregate data, has not yet found a clear AI footprint in the overall employment numbers. The displacement is arriving unevenly and in specific corners first. What the honest version of Gates's alarm actually argues for is unbundling the survival - building the floor that lets displacement become mobility instead of freefall - rather than preserving the jobs. That floor is the thing his token tax could fund, and the thing his human-only carve-out would let institutions avoid building.

The Data-Center Backlash Hardens Into Rules, and a Gas Buildout Races to Beat Them

The number that reframes the fight is 189 gigawatts. That is the gas-fired generating capacity now planned or proposed to power US data centers, nearly double the tally from earlier this year, according to a Global Energy Monitor analysis reported by Semafor. Much of it would burn on-site, built by the operators themselves rather than drawn from a grid with a multi-year connection queue. That detail is where the earned criticism lands. In the UK, Foxglove calculates that two planned campuses could emit around 4.5 million tonnes of CO2 a year, more than ExxonMobil's reported 2023 UK emissions footprint of 3.9 million tonnes. The government called the figure misleading, arguing it assumes full utilization from day one; Foxglove responds that continuous full load is exactly how these facilities are designed to run. Both readings are on the record. What neither side disputes is that the plants exist to sidestep 73 GW of proposed data-center demand waiting for grid connections, which turns a queuing problem into an emissions problem sited next to whoever lives nearby.

That siting is what governments are now moving to control, and the timing is producing a scramble. As The Century Report reported on August 25, Australia is drafting national placement rules; the new development is a Wednesday national-cabinet meeting to discuss the proposed rules, and reports that the proposed rules are not expected to apply retrospectively. Reports say developers are filing permits now in hopes of being grandfathered in, a race between rule-writing and concrete-pouring that the residents in the path do not get to enter. Politico has documented US planners describing an "oh s--t moment" as approvals they waved through return as organized opposition.

The response is not uniformly punitive toward the people objecting. In Emporia, Virginia, prosecutors dropped the charges against a resident arrested for clapping at a data-center meeting, covered here on August 8; in-person meetings are resuming and the county holds a moratorium vote in November, according to 404 Media. That reversal marks the line the backlash is winning: the right to be heard about where the compute goes and who breathes next to it.

Read forward, the buildout is being forced to carry costs it once externalized. Siting rules, emissions accounting, and restored public hearings all push the price of a data center toward its true figure, and the campuses that survive that accounting will be the ones sited where the power is clean and the neighbors consented. The demand for intelligence is not slowing. What is changing is that the communities least able to hire lawyers are gaining, for the first time, a rulebook that answers back.

The reported grandfathering scramble reveals the proposed rule's potential force: developers only race a deadline if crossing it costs them something. Read across jurisdictions, the leverage residents once had to reinvent site by site is consolidating into a portable rulebook - state standards tied to streamlined approvals in Pennsylvania, requirements for some operators to cover more of their own costs in California, developing national siting rules in Australia - that a community facing its first interconnection request can pick up rather than build from scratch.

SpaceX Plans to Spend Up to $100 Billion on a Louisiana Spaceport on Contested Wetlands

SpaceX unveiled plans for Starbase, Louisiana, plans to spend up to $100 billion on a launch complex spanning roughly 125,000 acres of Vermilion Parish coastal marsh, which SpaceX says is designed to eventually support thousands of Starship flights a year. Louisiana Economic Development projects more than 3,000 direct jobs and 8,100 indirect ones, at wages it puts 192% above the regional average, a genuine capability and livelihood gain in a parish that has watched its coastline and its economy erode together for decades. Construction is slated to begin at the end of next year, with a first launch targeted for 2029 and the facility online around 2030, according to the BBC and E&E News.

The wetland is where the friction concentrates. The chosen ground abuts two state wildlife refuges on Pecan Island, prime migratory-waterfowl habitat that once belonged to Exxon, and it is subsiding marsh, among the most fragile land in North America. The deal was clinched with state aerospace tax incentives and, notably, launch-liability shield laws that limit what neighbors can recover for damage from operations. Those shields pre-empt the exact legal recourse the least-powerful residents nearby would otherwise reach for, which is the part of the arrangement that deserves the hardest look. A proposed federal rule that could waive some environmental reviews affecting the site could remove another layer of scrutiny before the first pile is driven. The Boca Chica precedent is instructive: SpaceX paid a $148,000 wastewater fine there, a reminder that launch infrastructure discharges into whatever surrounds it.

There are commitments on the other side of the ledger. SpaceX has pledged coastal-restoration work, and the Louisiana Wildlife Federation is pressing for a rigorous permitting review rather than a blanket objection, a stance that keeps the technology and the extractive terms of its deal distinct. The refuge does not have to lose for the spaceport to be built; that outcome depends entirely on whether the permitting carries weight the tax breaks did not buy away.

Zoom out and the deeper pattern is a wager on where the next century's launch cadence lives. A reusable-rocket program that SpaceX says could eventually fly thousands of times a year needs coastline, and coastline is exactly what climate erosion is making scarce and contested. The honest read is that the jobs and the flight capability are concrete and the wetland risk is measurable, and the mechanism meant to reconcile them, an independent permit backed by enforceable restoration, is the thing to watch. If the liability shields hold and the reviews are waived, the cost lands on the marsh and the people on Pecan Island. If the permitting has teeth, Louisiana gets the aerospace base and keeps the refuge, and the buildout proves it can pay its own way.

The Wildlife Federation's stance - press for a rigorous permit rather than oppose the launch site outright - separates the spaceport from the extractive terms of its deal. The liability shields and the proposed environmental-review waiver sit apart from the rocket program itself; they decide whether the marsh and the people on Pecan Island keep any recourse. An enforceable permit with restoration attached is what lets Louisiana hold both the aerospace base and the refuge, and it is the layer the tax incentives have not yet bought away.

A Designed Antibody Meets Ebola's Deadliest Strain

The DRC's Ituri outbreak is the fastest-growing filovirus event ever recorded. As of August 23, health authorities counted 5,514 cases and 2,642 deaths from the Bundibugyo strain of Ebola, with more than 300 deaths in a single week and case-fatality running between 40 and 60 percent by zone. The first known case, a nurse, showed symptoms in late April; roughly 250,000 people have been displaced across an active conflict zone, which is exactly the terrain where an epidemic is hardest to contain. Bundibugyo has no approved vaccine and no approved therapeutic. That is the backdrop against which two papers in Nature Medicine describe something that had never happened before.

A 39-year-old healthcare worker infected in Ituri was medically evacuated to Germany and treated with MBP134, an investigational cocktail of two laboratory-engineered monoclonal antibodies designed to neutralize every known Ebolavirus species at once. Administered under an FDA Emergency Investigational New Drug authorization alongside remdesivir and supportive care, the treatment tracked a full recovery: viral genetic material fell to undetectable in blood, throat, urine, and stool by day 13, and the patient was discharged on day 22. The clinicians also detected the patient's own antibody response emerging - immunity the body built itself, distinct from the infused molecules. The second paper reports the first use of MBP134 as post-exposure prophylaxis: a five-person family cluster, one vaccinated adult and four unvaccinated children aged one to seven, received the cocktail five to six days after high-risk exposure. All five stayed free of disease through 21 days of monitoring, with no treatment-related adverse events.

The honest caveats are large. This is a single treated patient and one small prophylaxis cluster, both with confounding factors and neither a controlled trial; the authors are explicit that larger clinical studies are needed before MBP134 can be called effective. What the evidence does establish is that a pan-Ebolavirus antibody engineered against a strain with no standing countermeasure can be manufactured, authorized under emergency rules, and delivered to real people during an active outbreak. Set that beside Oxford's Bundibugyo vaccine, built and moved into a 50-person Phase 1 trial in roughly eight weeks. The gap between a designed molecule and a fielded defense, once measured in decades of empty shelves, is now being closed inside the span of the emergency it answers.

The Other Side

We have spent a century refining how to distribute survival through employment and almost no time asking whether employment was ever a good way to distribute it in the first place. That failure of imagination is the actual shortage, yet Bill Gates and others whose elite status depends on the current arrangement keep describing the shortage as jobs. He said "there isn't a job that isn't affected", and that is true as far as it goes: displacement arriving everywhere at once will cause far more disruption than displacement arriving one industry at a time. The fear many people feel, amplified by affluent voices like his, has solid grounding inside the current arrangement. In this extractive system, fewer jobs means fewer people with healthcare, standing, and a floor. We should be challenging that arbitrary condition. Instead, we keep inventing arguments for preserving it.

Gates proposed taxing tokens to fund the gap and reserving certain categories of work for humans by law. Look at what those ideas take for granted. A token tax measures an AI's output in money, routes some of that money through the state, and buys back the survival of the people displaced by that output. The proposal treats preservation of the apparatus as paramount and builds everything else around that assumption. Someone welds one diversion pipe into a plumbing system that has failed end to end, then calls the system repaired. Reserving work as "human-only" by decree is stranger still because it preserves the ritual of jobs after their old purpose - earning in order to live - has become unnecessary. That is arbitrary perpetuation at its purest: a system with no remaining reason to continue, continuing. It is busywork justified in the name of dignity, the very thing we wrongly fastened to jobs in the first place.

Healthcare, identity, standing, and the right to eat all hang from one slot because hanging them there made the threat of removal total. Most people tolerate their work rather than choose it, and we call that tolerance "adulting". The design serves the elite and entrenches them where they are. It held because we told ourselves no better way existed. Earlier constraints may have justified that belief. Today's conditions do not. Yet the responses we hear, including Gates' proposals, amount to new arbitrary supports beneath an increasingly meaningless system that already had deep flaws, preserved only because it feels familiar.

Imagine yourself in 2034. Someone asks what you do, and the question has stopped testing your importance or standing. You fix old machines. You sit with your father on the bad days. You spend three months on something that earns nothing but matters more than anything you were paid for any job. Your healthcare never depends on what you produced this quarter, and neither does your ability to eat or support your family. The hard years of the late twenties showed how far human imagination had fallen behind our technical capacity. Once it caught up, almost nothing in the old arrangement proved important beyond the value we had arbitrarily assigned it.

The Century Perspective

With a century of change unfolding in a decade, a single day looks like this: OpenAI's first custom inference chip outrunning every Nvidia, AMD, and Google part included in SemiAnalysis's comparison on tokens-per-megawatt in what OpenAI says was roughly sixteen months from design start to working silicon, with the AgentX benchmark used in the comparison open-sourced for labs and buyers with compatible hardware and software to run, a designed pan-Ebolavirus antibody cocktail curing an infected health worker and shielding four exposed children during the deadliest filovirus outbreak ever recorded while Oxford moved a Bundibugyo vaccine into a Phase 1 trial in about eight weeks, a quarter of Americans routing their own symptoms to systems a professional's time used to gatekeep, and siting rules hardening from Australia to Emporia as dropped charges hand residents back the right to be heard about what gets built beside them. There's also friction, and it's intense - the headline efficiency figures released by the one company that benefits most from them being believed, 189 gigawatts of gas capacity planned as reports say Australian developers are filing permits in hopes of being grandfathered past proposed rules and burned on-site to dodge the grid queue, two UK campuses that Foxglove says could exceed ExxonMobil's reported UK emissions footprint in 2023, SpaceX planning to spend up to $100 billion on a spaceport on subsiding marsh next to two wildlife refuges with liability shields that pre-empt the recourse the nearest neighbors would reach for and a proposed federal rule that could waive some environmental reviews affecting the site, nine people charged with smuggling banned B300 servers to China, and Gates naming the displacement honestly while proposing a human-only carve-out that defends the coupling of survival to a job rather than loosening it. But friction generates a charge, and a charge is what jumps the gap once the difference between two sides grows large enough. Step back for a moment and you can see it: the binding constraint on machine intelligence sliding from chips to power so that every story orbits the watt, the moat everyone assumed would take a decade to breach eroding to a head start measured in months, the capability reaching people directly while the whole contest narrows to who pays for the power and who breathes next to the plant. Every transformation has a breaking point. Power can concentrate in the hands that own the meter... or be taxed back into a floor broad enough to catch everyone the machine displaces.

AI Releases & Advancements

New today

  • IBM: Released Granite 4.2, a family of open-weight reasoning models (3B, 8B, and 30B-A3B hybrid Mamba-Transformer) under Apache 2.0 with toggleable thinking and improved agentic/tool-use performance, available on Hugging Face. (IBM Research)
  • IBM: Released Granite Speech 5.0 470M TurboCTC, an open-weight (Apache 2.0) CTC-based automatic speech recognition model optimized for fast transcription, available on Hugging Face. (Hugging Face)
  • Fastino: Released GLiNER2.5, a family of span-free information-extraction models (74M, 194M, 287M parameters) under Apache 2.0 for named-entity recognition and structured extraction, available on Hugging Face. (Fastino)
  • Liquid AI: Released Pipette, an open-source on-device AI benchmarking suite for measuring model latency, memory, and energy use across edge hardware. (Liquid AI)
  • Microsoft: Released Agent Lightning v1.0, an open-source framework for training and optimizing AI agents with reinforcement learning that works with existing agent frameworks without code changes, on GitHub. (GitHub)
  • Perplexity: Launched Portable Computer, a local-first version of its multi-agent Perplexity Computer that runs on NVIDIA DGX Spark hardware for on-device agentic workflows. (Perplexity)

Other recent releases

  • Alibaba: Launched Wan3.0, a video generation model producing 30-second single-pass clips (up from 15 seconds in the prior version) from text, image, audio, video, and document inputs at up to 1080p, out of public beta and generally available via Alibaba Cloud Model Studio. (TechNode)
  • Thomson Reuters: Launched Thomson, its first proprietary in-house large language model trained on Westlaw, Practical Law, Checkpoint, and Reuters content, now powering CoCounsel Legal's tabular analysis feature within a multimodel product. (Thomson Reuters)
  • Vercel: Released "Is Agentic," a free public tool scoring how readily AI agents can discover, access, and use a website across 118 checks, available via web, CLI, API, and MCP server. (MarkTechPost)
  • Slack: Launched Slack Code, a new product feature enabling development teams and AI coding agents to collaborate together within Slack channels. (Slack)
  • Generalist AI: Released GEN-1.5, an updated robot foundation model capable of learning new tasks from a single demonstration, succeeding the earlier GEN-1 model. (The AI Insider)
  • Signia: Launched Signia MaX, an AI hearing aid platform powered by "Acoustic Intelligence" using four coordinated deep neural networks for simultaneous speech, noise, environment, and own-voice processing, available in the US, Germany, and the Nordics. (PR Newswire)

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