Week in Tech: AI Chips Dominate Data Centers, CAR-T Goes Viral Against Autoimmune Disease, and Why Nobody Signed That Executive Order

The publishing world has talked itself hoarse about AI. It is still the most consequential technology story of our time — not because the surface is still novel, but because beneath every plausible headline there is another, sharper one worth making. In this week's edition we cover five genuinely moving developments across AI infrastructure and governance, model proliferation and product strategy, automotive electrification and autonomy, and biotechnology — running from chip-market dominance to gene-therapy applications to a surprise de-extinction breakthrough that may accidentally advance human fertility science. Here is what happened this week that actually mattered.

The Great AI Chip Race Accelerates

There is a line of thinking about the AI chip market that says Nvidia is simply too far ahead to be caught. That line of thinking just got harder to make after Q1 FY2027 earnings. Nvidia posted record overall revenue of $81.6 billion and record data-center revenue of $75.2 billion, up 92 percent year over year. The company remains the unelected governor of the AI infrastructure world, and that power is anything but vacated.

The magnitude of these numbers requires a moment of orientation. A single quarter at Nvidia's data-center division generated more revenue than nearly every Fortune 500 company posted in an entire fiscal year. The reason is simple: every meaningful AI model in existence — GPT, Claude, Gemini — runs through Nvidia GPUs, and as those models grow larger, the compute bills grow faster. Demand is currently bounded by how fast Nvidia can ship silicon, not by how fast customers want to buy it.

But the story beyond Nvidia's headline is just as interesting. Microsoft commissioned its own custom silicon — the Maia 200 — and not only uses it internally, it is now leasing it out to third-party AI companies as well. Anthropic, one of the most well-funded and strategically important AI companies outside of OpenAI, recently confirmed it is in early talks to rent Azure infrastructure running on Microsoft's Maia chips. This is a serious development: Anthropic already operates under one of the most celebrated compute agreements in Silicon Valley history (the now-legendary $15 billion per year SpaceX deal), yet that alone hasn't been enough to satisfy its capacity appetite. Adding Microsoft Azure as a supplementary provider means Claude will have more training capacity during the most competitive phases of model development.

Why Custom Chips Matter (Beyond Marketing)

The move toward custom AI silicon by hyperscalers is not merely a hedge against Nvidia pricing. It is a strategic shift with profound implications for the future of the industry. When Microsoft optimizes a chip specifically for the kinds of workloads Claude actually runs — inference scaling, fine-tuning, and retrieval-augmented generation — the efficiency gains can be significant. Maia 200 is not training hardware; it is designed for running existing models more cheaply and at lower latency. That is precisely what Anthropic needs as Claude expands from chat to coding to enterprise deployment.

Anthropic's simultaneous expansion of Azure usage signals what industry insiders have long suspected: even the most generously financed AI companies are forced into multi-cloud strategies, not because they want to diversify their vendors, but because the AI arms race demands more capacity than any single provider can credibly offer within a relevant timeframe. This pressure is the defining structural constraint of the industry right now.

AI at the Application Layer: PowerPoint, Policy, and Paranoia

ChatGPT Is Now Inside PowerPoint — For Real

OpenAI quietly shipped an integration that changes what people mean when they say "AI in the office": ChatGPT is now available as a native sidebar inside Microsoft PowerPoint, across Business, Enterprise, Edu, and even free ChatGPT tiers. Users can generate presentations, reformat slides, and summarize document fleets — all without leaving the app window. This follows the earlier Excel integration and follows the same pattern: OpenAI is embedding its language model into the workflow software people actually use for their jobs.

The significance is not the feature itself — anyone with a GPU and a developer key can produce a slide deck. The significance is the distribution mechanism. PowerPoint has 1.5 billion installs roughly. OpenAI does not need to reach those installs directly; it needs to convince Microsoft that its model is good enough to be the default AI behind every PowerPoint action. By shipping at the free tier, OpenAI is training the market to expect AI slides at no marginal cost, and once that expectation sets in, upgrading to Pro becomes frictionless. That is not a product strategy. It is a moat.

The Executive Order That Never Was

On Thursday, May 21, 2026, Donald Trump was scheduled to sign an executive order that would have established the most ambitious government framework for AI governance in American history. He cancelled it at the last minute. The ostensible reason, according to Politico, was that he "didn't like certain aspects of it" — diplomatic framing that presumes the document had already settled on its contents. That is rarely how last-minute EO cancellations work.

What we can say confidently: Trump did reference China explicitly. "We're leading China. We're leading everybody, and I don't want to do anything that's going to get in the way of that." This suggests the executive order contained provisions — likely around model export controls, compute licensing, or federal procurement standards — that might constrain American AI companies in ways relevant to international competitive positioning. An executive order that codified AI safety standards while simultaneously handicapping frontier model developers in the US relative to unregulated Chinese competitors is the kind of document that makes leaders choose pragmatism over principles.

In practical terms, the US has been operating without comprehensive AI federal governance since the last administration's executive orders expired under political transition. AI regulation in the US is currently patchwork: state-level bills (California, Colorado), industry self-regulation, and the informal guardrails exercised by major AI companies who prefer not to blow up the market before it matures. Until this EO was started, the process of writing it was arguably the most important AI governance conversation happening in Washington in 2026. Its cancellation is the most important single non-development of the week.

Inside OpenAI: A Split Happening in Public

Aleksander Madry, one of OpenAI's most prominent safety executives — formally the "head of preparedness" before being reassigned last summer — announced his departure on Thursday. He is leaving to work on "something new, centered on AI's impact on the economy." That phrasing is notable. A departure from OpenAI that centers "AI's impact on the economy" rather than "AI safety" or "AI alignment" implies a field of action that extends beyond the technical purity of alignment research and into where AI is actually changing labor markets, capital allocation, and competitive dynamics. Madry's decision to leave OpenAI — arguably the place with the most resources for any AI-related project — and choose a topic that is simultaneously adjacent and outside the core research agenda is itself a signal that the AI safety conversation is fracturing into a more complex ecosystem.

This comes against the backdrop of the Musk v. Altman trial, in which OpenAI is defending itself against Elon Musk's claim that Elon is owed equity in the company, and the trial has produced surface-level chaos in the form of statue-of-ass evidence submissions. Ilya Sutskever, co-founder of OpenAI and once one of the most reclusive and respected figures in AI, testified that he stood by his role in Sam Altman's 2023 ouster and stated, "I didn't want it to be destroyed." These public读取s of OpenAI's internal governance and philosophy are not chemistry drama — they reveal genuine fault lines over control of AI development at the highest level.

Google DeepMind, Gemini, and the Quest for Agentic AI

Google DeepMind CEO Demis Hassabis made a bold claim at this year's Google I/O keynote: Gemini was positioned as a more reliable and capable reasoning engine than anything the market had yet validated at scale. The claim was bold enough to attract specific pushback from analysts who noted that many of the flashy Gemini capabilities demonstrated at I/O had not yet shipped to users. This is a recurring pattern: Google demonstrates cutting-edge AI capabilities in keynote form and the execution cadence lags behind.

The latest delivered item: CapCut is coming to Gemini. CapCut, the mobile-first video editing tool backed by ByteDance, announced it will let users edit images and videos directly inside the Gemini application — no app-switching required. The stated philosophy from CapCut is that "the future of creation will be more conversational, intuitive, and intelligently integrated across tools and experiences." That is a fair summary of where Google believes its competitive wedge is: not just smarter models, but smarter model *integration*. If the model can touch the right file, run the right operation, and return a result without you leaving the workflow, the entire notion of a "product" begins to dissolve into a surface.

There is also the ongoing Intuit layoff — approximately 3,000 employees, or 17 percent of the workforce — which the company explicitly framed as a move to accelerate AI investment in its products. TurboTax and QuickBooks running on AI pipelines is not hypothetical; it is a stated near-term priority. This is the real mechanism of AI displacement: not entertaining chatbots, but pervasive AI rewriting core software workflows, layer by layer, company by company.

EVs and Autonomy: The Quiet Progression

The broader electric vehicle market is going through a classic technology adoption correction — too many manufacturers, too much supply pressure, and not enough battery-cost reduction to justify the subsidy-driven growth of the early 2020s. Yet the sector is not stalling; it is consolidating and making real engineering progress at the margins.

On self-driving technology, progress continues at pace. Waymo's robotaxi fleet continues to expand in previously underserved markets, and Tesla's FSD (Full Self-Driving) is advancing toward regulatory and technical milestones that, if cleared, could change the economics of ride-sharing in a way that makes conventional ownership look expensive. A fleet of electric robo-taxis operates at a substantially lower per-mile cost than human-driven ride-sharing only when autonomy is sufficiently reliable and the regulatory path for commercial deployment is clear. We are closer to that threshold today than at any prior point in the EV industry's history.

The bigger vehicle technology news is in battery chemistry. Solid-state battery development — which promises roughly double the energy density of current lithium-ion cells at lower fire risk — is approaching lab-to-production scaling. Toyota has been the most aggressive about timelines, with multiple iterations of solid-state cell prototypes appearing in testing vehicles. If commercialization hits before the end of the decade, the EV form factor could shift dramatically: smaller batteries, faster charging, and electric vehicles that genuinely compete with gasoline cars on range and convenience without tech-luxury pricing.

Biotech: CAR-T Leaps Beyond Cancer

The Autoimmune Reset That Changed a Life

At age 49, Jan Janisch-Hanzlik's multiple sclerosis was actively limiting her physical freedom and her sense of safety. She gave up her active nursing career for desk work. Frequent falls made carrying her grandchildren a source of anxiety. She had planned to move into a bigger home specifically in case she needed a full-time wheelchair. Conventional disease-modifying therapies were not improving her condition. So she phoned the University of Nebraska Medical Center every other month until they were willing to enroll her as the first patient in a CAR T cell therapy trial for multiple sclerosis.

She received the treatment in June 2025. CAR T cell therapy was originally designed and approved for use against blood cancers in 2017. The basic mechanism is relentlessly elegant: T cells — the immune system's primary hunter-killer cells — are extracted from the patient's blood, engineered to express a chimeric antigen receptor (CAR) that targets a specific molecular characteristic of the patient's disease (in cancer: one or more markers on malignant cells), and reinfused into the patient where they hunt and destroy their targets.

Today that same technology is being applied — in hundreds of clinical trials — to autoimmune diseases including multiple sclerosis, lupus, Graves' disease, and vasculitis. The logic is biochemically symmetrical: in autoimmune disease, the immune system is hunting cells labeled "self" when it should not. If CAR T cells can be engineered to destroy the specific immune cells responsible for attacking the self — rather than the entire immune system, which is what current immunosuppressive drugs do — the result would be a targeted immune reset rather than a wholesale system downregulation.

Janisch-Hanzlik knew the risks. She had two young grandchildren. Multiple sclerosis has a significant genetic component. She also had professional expertise as a nurse and was not making an uninformed decision. What's extraordinary about this case is that a therapy originally designed to kill cancer is now being tested across hundreds of autoimmune conditions, and the first-patient outcomes are generating cautious optimism in clinical researchers who have seen previous "breakthrough" immune therapies fail for autoimmune populations.

The Artificial Egg That Accidentally Changed Developmental Biology

Colossal Biosciences, a company best known for its de-extinction ambitions — woolly mammoth-like elephants, thylacine cloning — may have accidentally made the most important developmental biology tool in a generation. While attempting to produce a chicken egg capable of housing de-extinction hosts, the company's researchers discovered that they could generate an egg from a hen without standard fertilization. The result: an egg without the genetic contribution of a rooster, yet with the biology of a viable egg — yolk membrane, extracellular protein framework, and the whole reproductive geometry a developing embryo would require.

The primary ambition of this artificial egg work is to support Colossal's broader de-extinction program. The secondary, and from the standpoint of developmental biology far more significant, application is in the study of how fertilized eggs develop into complex organisms without the confounding factor of reproductive heterozygosity. Researchers studying embryogenesis — how a single fertilized cell becomes a differentiated human being — have been hamstrung by the fact that every embryological model organism has its own DNA's fingerprints animating the process. An artificial egg offers a blank canvas that may allow researchers to isolate individual genetic and epigenetic variables with unprecedented precision.

AI and Science: Two New Tools That Actually Work

Ars Technica reported that two AI-based science assistants have independently demonstrated drug-retargeting capability at a useful accuracy level. This is not hype-laden AI science-fiction territory — both tools were benchmarked against real repurposing tasks and completed them. One of the two went further and analyzed the experimental data internally, rather than just generating hypotheses for a human scientist to test.

The significance here is that repurposing existing drugs to treat new conditions is one of the more cost-efficient paths in pharmaceutical development. Developing a new drug from scratch — identification, preclinical, Phase I–III trials, regulatory approval — averages a billion dollars and a decade. Repurposing an existing, already-approved drug for a new indication bypasses most of that cost. If AI tools can reliably identify repurposing candidates — cost-effectively, reproducibly, and with strong mechanistic justifications — the pharmaceutical industry's risk profile and its calendar of new therapies both shift meaningfully.

The Pattern Across All Five Stories

Taking a step back, each of these stories shares a common thread: the cleanest developments in 2026 are happening at the junction of disciplines, not within them. The AI chips story lives at the intersection of semiconductor engineering and AI infrastructure economics. The Microsoft Maia-Anthropic partnership is a story about software-AI economics, not just chips. The CAR T cell autoimmune story is an oncology technology applied to immunology with unexpected results. The artificial egg story is a de-extinction experiment generating a tool with developmental biological implications it didn't anticipate. The AI drug repurposing paper is a computational tool applied to pharmaceutical science at a scale that wasn't possible before modern transformer architectures.

What to Watch Going Forward

In AI, the decisive question of the next 12 months is not which model is smarter on benchmarks — it is which AI model can embed itself into the highest-value workflow on the largest installed base. PowerPoint's adoption matters more than any academic paper because it touches the day-to-day work of hundreds of millions of people. The companies that win will be the ones that integrate, not the ones that impress.

In biotech, the next 12 months of CAR T autoimmune trials will determine whether Janisch-Hanzlik's story becomes a category of treatment or remains an inspiring but isolated result. CAR T in cancer worked because cancer cells have selectable surface markers. Autoimmune conditions are more heterogeneous at the cell level, meaning the engineering challenge of the CAR receptor is harder and more bespoke.

In electric vehicles, the battery race is the right one to follow. Solid-state progress has been incremental by design — the materials are expensive and the manufacturing tolerances are unforgiving. But when a breakthrough on the manufacturing side arrives, the product category changes in a single generation.

Technology moves in sweeps rather than linear steps. The sweep happening now — AI applied at the user-applications layer, AI and biotech deepening their integration, and compute infrastructure consolidating around a handful of truly dominant players — is one that will define the technological landscape for the balance of the decade.

---

This post is part of our weekly deep-dive into the technologies that are actually changing the world — excluding the noise. Subscribe for the full analysis every Friday.