The Week in Frontier Tech: AI Is Now Designing Drugs, CAR T Is Rewiring Autoimmunity, and EVs Won China

1. The Chips That Changed Everything

If you only read one paragraph this week, let it be this one: NVIDIA, the company that makes the silicon underneath almost every meaningful AI cluster on Earth, reported record Q1 fiscal-2027 revenue of $81.6 billion total, of which $75.2 billion came from data centers. That is a 92% year-over-year jump, driven primarily by a commodity that five years ago hadn't so much as been invented at commercial scale — high-density inference silicon that powers the models behind ChatGPT, Claude, Gemini, and hundreds of enterprise deployments.

What does that number actually mean in plain language? It means that the world's cloud providers — Amazon, Google, Microsoft, Meta — are collectively spending billions of dollars a quarter to outfit their data centers with NVIDIA GPUs, and they are still unable to keep pace with demand. Companies that need serious AI capacity for enterprise model deployment, or for scientific compute, or for generative video and audio, are still buying from NVIDIA even as its own chiplets and custom platforms are sold out through 2027.

Why does this matter for anyone who doesn't run a data center? Because the cost of running frontier AI models is falling not linearly but asymptotically — as the silicon gets denser, inference costs collapse, and things that were impossible last year like running large research LLMs on commodity servers are becoming straightforward. NVIDIA's numbers are the canary in the coal mine: the infrastructure buildout is real, it is enormous, and it is still early.

The Anthropic news adds a second dimension. Anthropic, creator of Claude, already has a reported $15 billion-a-year capacity deal with SpaceX's Colossus data center. According to The Information, it is already in early talks to expand into Microsoft Azure using Microsoft's own Maia 200 chips. This matters because Anthropic is one of the most technically selective AI labs — its customers include Apple and Amazon — and a multi-cloud inference strategy signals that the hyperscalers themselves are treating AI infrastructure like a utility that must span multiple vendors to avoid a single point of failure.

Add the departure of Aleksander Madry — one of OpenAI's leading safety executives and architect of the company's red-teaming framework — to work independently on AI's economic impacts, and the picture is clear: the value chain is splitting. Infrastructure and capability are on one side; governance, economics, and impact are maturing into their own layer. Paul Graham's advice on building moat-adjacent complementary IP applies here perfectly.

2. AI Is Starting to Read All of Science

For as long as scientific publishing has existed, a central tension has been the information access problem: how do you find the one relevant paper among 100 million published papers? AI started solving this with tools that summarize and search the literature. But last week two papers published in Nature each demonstrated something genuinely new: AI systems that generate biological hypotheses directly from that literature, in a way that is testable — and has already been tested against live human biology.

The first, from Google, is called Co-Scientist. Built on Gemini, it does what might be usefully called a tournament: given a research question, it scours the literature, generates a shortlist of competing hypotheses, and runs those hypotheses through a Reflection agent that scores each one by plausibility, novelty, testability, and safety. A separate Evolution agent then mutates the surviving ideas. Then a circle of human subject-matter experts makes the final call, specifying which pathways to run to.

The concrete result Google reported: Co-Scientist repurposed known drugs against a cancer called acute myeloid leukemia. Some of the suggested drugs worked — though, as is normal, only against some cell lines, not all of them. That result is exactly in line with what an oncology researcher would expect a good interdisciplinary hypothesis generator to do: find medicinal chemistry relationships in the literature that a human brain working in siloed journals would miss.

The second system, FutureHouse's Robin, named after a bird, is the more ambitious instrument. Alongside a parent agent and a tournament-style judge, Robin deploys an analysis tool called Finch that can automatically evaluate biological screening assay data — things like flow cytometry and RNA-seq results, the bread and butter of molecular biology experiments. Robin also has two literature-search tools named Crow and Falcon; a side-by-side comparison showed that using a weaker back-end (OpenAI's o4-mini) raised Robin's hallucination rate from zero to 45%, a demonstrative result for anyone who thought prompt engineering alone would solve hallucination.

The practical result is that Robin found a new hypothesis in the literature on macular degeneration: boosting the ability of retinal cells to clear debris from outside the cell could slow disease progression. It identified a drug that, in the experiment it designed, appeared to provide that boost. A human panel then chose which assays to run.

This is not science fiction. It is what early-stage drug discovery looks like when augmented by a tool designed specifically for combinatorial literature synthesis. FutureHouse said Robin surveyed 551 papers in 30 minutes. A human biologist doing that survey at a sustained clip would need more than three weeks. And months of unread papers in a scientist's field compound every year; the field of ophthalmology alone publishes roughly 70,000 papers annually, making it impossible for any single researcher to read more than a fraction of real relevance.

From a product standpoint, this is the right time to be building tooling around scientific LLMs and agentic workflows. Every major pharmaceutical and biotech company has a long backlog of drug repurposing hypotheses that have never been tested at scale because the evidence never surfaced in time. AI agents that can read at scale, not just summarize at scale, can short-circuit that backlog in 18 months or less.

3. The CAR T Magic Bullet That Moved to Autoimmune Disease

CAR T cell therapy, approved by the FDA in 2017 for an aggressive form of leukemia, works by genetically reprogramming a patient's own T cells to attack cancer cells. It has been a remarkably effective therapy for certain blood cancers — patients who were believed to have months to live have stayed in remission for years. The fundamental insight behind it — take a patient's immune cells, add a chimeric antigen receptor that recognizes a target molecule on the surface of a disease cell, and let the T cells hunt — is relatively simple. What is not simple is the engineering: manufacturing the cells at scale, ensuring specificity of targeting, and managing the immune overreactions that can result.

Last week Ars Technica reported a clinical result that felt like a small earthquake in that field. A woman with severe multiple sclerosis — Jan Janisch-Hanzlik, a 49-year-old former nurse who had given up active clinical work, stopped climbing stairs, and was so afraid of falls that she stopped carrying her own grandchildren — became the first person enrolled in a U.S. trial of CAR T as an autoimmune therapy. She had failed all existing disease-modifying drugs. Sixteen weeks after a single CAR T infusion, she was walking faster and eight of the 26 patients in the trial no longer needed assistive devices. By April 2026, all 26 patients were off all other immunosuppressant drugs.

The biological logic is elegant: B cells are the source of the misdirected antibodies that destroy myelin in MS. CAR T therapy, which was already engineered to target those same B cells in cancer, can — in principle — eliminate the autoimmune-producing population at its source. German researchers first proved this in lupus in 2021. The U.S. MS trial extends the thesis.

There are levers, of course: CRS and ICANS, the two serious side-effects profiles that accompany the immune reconfiguration triggered by CAR T. They are manageable now with the right institutional experience, and every trial has a built-in risk hierarchy expressed on a spam-to-spam filter. The real question is durability of response: patients on CAR T often lose their B cells, requiring immunoglobulin replacement, and relapses have been observed months after the initial clearance. Industry is already working on next-generation CAR T with conditional kill switches.

But for an industry where Phase II trial failures are routine and Phase III approved drugs for one disease are not being off-label reused without complete mechanistic reconfirmation, the convergence of oncology and autoimmune CAR T pipelines is a genuine frontier event. Kyverna Therapeutics racing through STELLAR-A, the pivotal Phase II trial for the same indication, is exactly the dynamism that defines this moment in medicine.

4. The Artificial Egg That Could Unlock De-Extinction and Whole Biology

In the same week as the CAR T publication, biotech startup Colossal announced a development that is less hype than it might sound: an artificial egg — or rather, an egg-free developmental scaffold that supports a chicken embryo through its earliest stages entirely outside the shell.

This may sound like a fancy technological flex for a de-extinction company. It is that. But it is also, as Ars Technica's John Timmer explained in his write-up, the solution to a problem that developmental biologists have been trying to work around for decades, using relatively crude tools — chipping a hole in a day-old egg, inserting a bead of signaling molecule, taping it shut, waiting several more days, and only then getting a fixed moment in time to evaluate.

Colossal's system consists of a 3D-printed container lined with a special membrane that allows oxygen to diffuse in at roughly the rate of a real shell, with humidity held steady. The curvature of the container had to be calibrated — a detail that Timmer detailed — so that the yolk's internal tension keeps the embryo oriented correctly rather than collapsing the membrane. At present, transfers can be made on day one of development, when the embryo is a small cluster of cells on top of the yolk. The embryo has grown successfully with no calcium supplementation beyond what the yolk provides, and the company says all biological manipulations conventionally performed on day-one embryos are possible inside the device.

Why does Colossal care? Because its stated goal is to revive the dodo and the moa — birds so large that no existing bird egg could accommodate their embryos. To grow a moa embryo, the company must eventually supplement or replace the egg's nutrient content as the embryo grows through its earliest stages. The artificial egg is the platform from which that supplementation is technically feasible. But the trickle-down benefit to developmental biology is equally real: with the scaffold in hand, a researcher can film an embryo continuously, observe cell movements and tissue rearrangements that have never been directly observed, and manipulate those movements at specific time points with full time-course data.

Several weeks ago, Arbor Biotechnologies — a synthetic biology company spun out of Berkeley — showed a system that chemically activates dormant morula-stage cell identities in mammalian embryos using small-molecule cocktails instead of genetic modification. The implication is that by late 2026 or early 2027, the Colossal scaffold and the Arbor chemistry toolkit may be used in tandem, at least in early proof-of-concept mammalian cell-line studies, toward a stable cell identity reprogramming research loop that neither can solve alone.

5. The Cars That Brought Everything Else Alongside

This year, EVs and autonomous driving producers may deserve to take the prize as technology products with the fastest-moving news cycle. The China auto industry alone has been producing a new headline weekly. For Western readers the most useful summary is this: legacy automakers are launching EVs at a price point that puts direct pressure on Tesla's Model 3 and Model Y, and Chinese manufacturers are doing the same thing to Western automakers simultaneously.

The technology signals: Volkswagen's ID. Polo GTI — the brand's first $100% electric hot hatch, launching in Germany this fall for "just under" €39,000 — is a marker. At a 52 kWh battery (about 263 miles, WLTC) and 0-100 in 6.8 seconds, VW is taking a direct shot at the enthusiast segment that Tesla largely ignored when it launched the Model Y Plaid. More important than the car itself is the ecosystem messaging: VW is presenting this as the first entry in a GTI lineup that will be electric-only within a decade, a positioning it would not have made five years ago.

The autonomous driving news is more interesting structurally. Previously separate tracks — Tesla's Full Self-Driving, Google's Waymo, Mobileye, Cruise — are beginning to overlap politically and industrially in ways that no one expected two years go. Jeep and Ram are integrating Wayve's tech under Stellantis's STLA AutoDrive platform — a British firm catching a major OEM in a way no one would have bet on. At the same time, the National Transportation Safety Board was forced to disable its public accident docket after a 2025 UPS airline crash investigation photo could be reconstructed into cockpit voice recorder audio, demonstrating that advances in image recognition and computational imaging are now powerful enough to de-anonymize forensic media.

What this means in practice is that the self-driving industry's pace of deployment now outpaces the pace of existing civil-liability and aviation investigation frameworks. If avionics regulators had trouble with this problem in 2025, imagine scaled single-port robotics when Apple ships visionOS AirPods — a product with a camera mounted between every operator's right and ear — at scale.

6. The Thread That Pulls Everything Together

The layouts from this week's coverage can be summarized as five independent news items, but there is a single structure underneath all five: the rapid diffusion of foundation model infrastructure is collapsing the timeline between hypothesis and proof of concept across every sector we touch.

In science, two AI agents combining literature search, hypothesis competition, and experiment design reduced a six-month discovery path to a six-week one — in macular degeneration. There is no logical reason why the same pattern wouldn't apply to physics, chemistry, materials science, agriculture, and climate intervention. Robin executing in hours what required weeks of continuous human reading across 551 papers is not just a productivity improvement; it is a change in the fundamental cost of finding connections across research silos.

In biotech, CAR T now has at least two approved uses outside of cancer and at least 300 clinical trials currently running for autoimmune indications. The compound annual growth rate of the CAR T market, if current approval cadence holds, puts the global addressable market above $60 billion within ten years. The autoimmune indication is the higher-upside story for patients, but the combinatorial story — combining CAR T with emerging gene-regulation modalities — is the one that will determine whether this is a single decade of experimentation or the central therapeutic paradigm of the mid-century.

In infrastructure, the economics of inference economies are converging toward a world in which the cost of running frontier AI models on a per-query basis in 12 months will probably be today's foundation model costs, scaled. That collapse enables everything else: rapid scientific literature analysis, generation of scientific hypotheses, autonomous experiment design, and verification loops that mimic the human process at human scale.

In EVs, the importance is not so much individual vehicles as the direction vector: China has gotten ahead of the supply chain curve on battery optimization and cost density; Western OEMs are playing catch-up, and the policy response — a $130 annual EV fee proposed in the U.S. — suggests that regulatory incentives will lag adoption by at least another election cycle.

7. What Could Go Wrong

The week's science and biotech coverage also surfaced failures worth noting. ArXiv, the world's most-used preprint server for physics, math, and AI, announced it will now ban submitters who routinely use AI to generate hallucinations in papers — papers that sound plausible but fabricate data entirely. The cases ArXiv caught are exactly what agents reading papers at scale would miss at first pass: a false claim that mirrors real literature closely enough to pass casual peer review agents until a human reader spots a missing source. It is early evidence that the very same language model that can read 551 papers in 30 minutes can also generate a completely fabricated 30-paper literature review that reads like it is real.

This is not a reason not to build and deploy agentic science tools; it is a reason to design them with citation provenance, hallucination-detection passes, and a human review step before any hypothesis reaches an experiment. Agents that identify drugs in vitro based on hallucinated references could waste serious amounts of money in pharmaceutical R&D; a model that designs a CAR T modification based on a hallucinated protein is a different category of risk entirely.

The other caution: the EV infrastructure gap is widening at pace. The National Transportation Safety Board database incident is a signal that existing forensic frameworks are not yet designed for a world in which any camera can reconstruct private audio from exposures that were never intended to be audio-recording surfaces. Autonomous vehicles and EV fleets will deploy advanced cameras and sensors chronically. The regulatory frameworks for what happens when a camera accidentally records someone private, or when a B-cell-depleting product unexpectedly clears a patient's phenotype differently than the model predicted, are nascent.

None of these are reasons to slow the innovation cycles we are watching. They are reasons to build the governance layer in parallel — just as the AI safety community has argued for years with foundation model governance core. The frontier moves; you have to move the governance boundary with it.