The Machines Are Working — Whether We’re Noticing Yet: AI Brains, Self-Driving Cars, and the Quiet Bio Revolution of May 2026

The AI Chip Wars Have a New Unexpected Front

Anthropic Looks Beyond Its SpaceX Deal

For much of 2025 and into 2026, the dominant AI infrastructure narrative revolved around who could raise enough capital and who could lay down the fiber to buy the most Nvidia GPUs. OpenAI, Meta, Google, and a handful of mutual-fund-style consortiums turned data-centre construction into a new form of industrial arms race. But a quietly reported detail in late May 2026 suggests the GPU-only model is fraying at the edges.

The Information reported that Anthropic — already crunching its way through a $15 billion-per-year deal with SpaceX’s Colossus supercomputer cluster — is holding early talks with Microsoft to rent servers running on Microsoft’s proprietary Maia’200 AI chips on its Azure cloud. The significance is hard to overstate: this would be one of the most direct acknowledgments yet that the AI model layer and the chip infrastructure layer are becoming dangerously intertwined. Microsoft has invested heavily in Maia and Cobalt as alternatives to Nvidia’s GPUs, and OpenAI has been a primary MoU customer up to now. If Anthropic formally migrates Azure workloads that direction, it gives Microsoft the narrative muscle to pitch Maia to every other LLM startup watching the capacity crunch.

Analysts have long noted that Microsoft’s arrangement with AI companies runs “hot and cold” — Microsoft is simultaneously an investor, infrastructure partner, customer, and platform host, making its long-term neutrality suspect. The Maia chip deal would shift that dynamic further. Maia’200 is not, like Nvidia’s Blackwell or Hopper, a training-first chip, but it is being architected specifically to run inference workloads efficiently at massive scale. With Anthropic’s Claude models fetching billions of API calls monthly, that’s exactly the sort of workload pattern chip buyers optimize for.

Nvidia’s $75.2 Billion Quarter

For now, that narrative shift hasn’t touched Nvidia’s bottom line. The company reported its first fiscal quarter of 2027 results in May, posting record overall revenue of $81.6 billion. Of that, $75.2 billion came from its data-centre division — a 92% year-over-year jump. The hyperscalers are still buying every chip Nvidia can ship. But if Microsoft’s homegrown chip strategy reaches critical mass — and if a growing number of model companies follow Anthropic’s rumoured lead — the five-year arc of compute infrastructure may start bending away from GPU monoculture sooner than most bulls expect.

AI Leaves the Chat Window and Steps Into Workflows

The Productivity Layer Arrives

While chip deals and APRU estimates are helpful sketches for investors, the AI story that most directly touches people every day is whether AI moves beyond the chat interface and into actual workflows. The clearest signal in May 2026 is that it is.

OpenAI announced a Microsoft PowerPoint add-on that lets users generate and edit entire slide decks from prompts, drawing on documents, images, and other source material. It sits as a sidebar inside PowerPoint, stuffed with enough integration ambition to make sense of a marketing deck in about the time it takes to pour a coffee. This follows a similar Excel and Google Sheets integration that shipped earlier in the year. Taken together, the integrations form the first genuinely coherent "AI workspace" layer — not just answering questions, but owning parts of the actual output.

Meanwhile, ByteDance’s CapCut announced image-and-video editing capabilities that will land directly inside Google’s Gemini app. The stated rationale is "the future of creation will be more conversational, intuitive, and intelligently integrated across tools." That phrasing sounds like marketing, but it points at an underlying structural shift: the next generation of AI-native tools aren’t building standalone products. They’re embedding directly into the apps people already reach for.

Andrej Karpathy Joins Anthropic

That embedding trend is also being turbocharged by the most quietly consequential talent move in AI in months: Andrej Karpathy, who led Tesla’s Autopilot effort and was a founding team member of OpenAI before leaving to build an AI-native school concept, is joining Anthropic on the R&D side. Karpathy is one of the most respected independent voices in deep learning — his microscopes-plus-LLMs approach to research combined with his experience leading large-scale ML infrastructure at Tesla means he brings a rare blend of engineering craft and research integrity to the table.

The move came a day after the news broke that Aleksander Madry, one of OpenAI’s top safety executives who was reassigned to an AI reasoning focus in 2024, had announced he’s leaving to work on AI’s economic impact. Together, the two departures open up massive questions about the inside culture at OpenAI and, see above, where exactly the talent gravity well for frontier AI research is settling. Anthropic has been steadily building that bank, and the Karpathy hire is the clearest confirmation yet.

The Worker Displacement Moment Is Real, and Starting Now

Not every story in the AI economy this month is a win for the industry. Intuit announced plans to lay off approximately 3,000 workers — around 17 percent of its global staff — with CEO Sasan Goodarzi framing the cuts as a way to double down on embedding AI across its products. The company builds TurboTax and QuickBooks, and its customers are millions of small businesses and individuals who will almost certainly interact with AI-enhanced versions of those tools in the coming months. That means direct job displacement with the product still in a training-and-rollout phase. It’s a useful, if chilling, signal: the productivity gains from AI-assisted software are now large enough to justify reorganizing an entire workforce around them. That math will be replicated across every industry that relies heavily on structured relational data.

The Car That Drives Itself Is Not Just Tesla’s Problem Anymore

Wayve and Stellantis: The Hands-Free Alliance

Autonomous driving has long been dominated by a handful of names: Tesla’s Full Self-Driving (still in supervised mode), Waymo’s robotaxis in Phoenix and San Francisco, and a handful of niche academic programs. May 2026 delivered two moves that expand the map in structurally different ways.

First, Stellantis — the transatlantic auto conglomerate that owns Jeep, Ram, Dodge, Peugeot, and Fiat — announced it will embed UK-based autonomous-driving software company Wayve’s technology into its STLA AutoDrive platform. The goal: hands-free, door-to-door supervised automated driving on both urban streets and highways. Think of it as Stellantis’ answer to Tesla’s FSD, but sourced from an outside partner rather than built in-house. Critically, Wayve has spent years building a vision-first approach that doesn’t depend on lidar arrays the way Waymo’s system does. Reducing the hardware bill of materials is what makes the economics of hands-free highway driving feasible at volume — and Stellantis’ fleet is measured in the tens of millions of vehicles.

This is a different kind of AV signal than Tesla’s approach. Instead of philosophy-first software that rolls out to customers as beta, this is a OEM-partnered software-at-scale integration. It matters because the mainstream car market turns on fleet economics, and Stellantis owns enough volume to make this work or break the financial viability of Wayve’s model. The deal is also likely to accelerate regulations in Europe and the UK around supervised autonomy, since a major OEM is now publicly flagging a commercial timeline.

Matrix LED and a More Boring Miracle

Second, and in some ways more quietly important, Audi announced it would bring its Matrix LED headlight technology to the United States. The system hasn’t changed much in technical architecture since it first shipped in Europe in 2013, but it’s the kind of feature that changes how cars interact with each other on the road: using cameras and real-time adaptive beam shaping to reduce glare for oncoming drivers while still providing maximum visibility. US NHTSA regulatory language had traditionally blocked this class of adaptive beam technology, but a 2022 rule change opened the door. Audi’s Q9 and SQ9 SUVs will be the first US units to carry it. Every other adaptive-beam technology founder in the US market will now have a litigation-friendly precedent to cite.

Mercedes-AMG’s Hollywood Autobahn Moment

While the policy pieces of the transport story are structural and important, we’d be remiss if we skipped the event that felt like the automotive equivalent of a Coen brothers film. Mercedes-AMG assembled 600 people, closed down Los Angeles’ famous 6th Street Bridge, and let the new EV four-door AMG GT launch with burnouts while Blink-182 played a 30-minute set. Brad Pitt was reportedly there. Jacob Elordi was apparently also there. Automotive marketing is long overdue for this level of "what is happening" sincerity, and the car itself — rumored north of 600 hp all-electric — is the kind of statement product that forces legacy critics to reckon with the electrification velocity of Stuttgart’s performance arm.

Pizza by Drone, Texas Edition

Finally, a story that feels both ordinary and strangely prescient: Israeli drone company Flytrex is building a manufacturing and maintenance facility in the Dallas-Fort Worth area specifically for drone deliveries after its Chinese imports ran headfirst into US regulatory friction. The company currently operates drone delivery for Uber Eats and DoorDash, and claims its larger drones can carry two large pizzas at once. The real story here is the decoupling of drone hardware from Chinese supply chains: with DJI effectively blocked in the US market, this is the kind of pivot that will shape a segment’s geography for the next five years. Opening thousands of deliveries across 60 DFW sites by mid-2027 is a real operational goal, not posture.

The Quiet Revolution in Biology Is Already Production-Mostly

Colossal’s Artificial Eggshell: De-Extinction Meets Developmental Biology

De-extinction company Colossal — whose stated ambition is to bring back the woolly mammoth and other lost creatures — announced a development that may prove more significant than its headline animals. They’ve built an artificial eggshell: a device that holds the contents of a chicken egg outside of the shell and allows the entire developmental process to continue normally, resulting in healthy hatchlings walking out of it.

The engineering problem was more subtle than it sounds. A developing embryo produces forces through the egg membrane that need precise structural tension to keep the embryo properly organized. If the membrane tension sags, the embryo crumples into a disorganized mess and fails to develop. Colossal’s device controls that curvature precisely, so the membrane tensioning works correctly throughout the entire incubation window. They transferred eggs within a day or two of being laid and had normal chicks walk away at the end.

The immediate application is for avian de-extinction work, specifically the effort to reconstruct the genome of the dodo and a handful of other extinct birds by assembling it in a chicken egg environment. But the secondary application is exactly the kind of unexpected cross-discipline utility that makes science investment interesting: an artificial egg that can be opened and closed without harming the embryo gives developmental biologists a continuous live view throughout the entire process of an organism’s formation. Extracting chicken embryos from shells and imaging them has historically been clumsy, destructive, and limited to one or two windows. This device removes one of the fundamental physical constraints of embryological research.

CAR T Cell Therapy for MS: Cancer’s Weapon Reborn Against Autoimmune Disease

Perhaps the most personally moving science story of the month is the emergence of CAR T cell therapy in autoimmune disease, specifically multiple sclerosis, where the first patient was successfully treated in 2025 at the University of Nebraska Medical Center.

CAR T cell therapy — approved by the FDA for certain blood cancers in 2017 — involves extracting a patient’s T cells, genetically engineering them to express a receptor that latches onto a specific target cell, and reinfusing them. In cancer, that target is malignant cells. In autoimmune disease, the target is the malfunctioning immune cells that are attacking the body’s own tissue. The promise is a hard reset of the patient’s immune system — replacing an overactive, self-damaging response with one that is (potentially) permanently corrected.

The concept is stunning in its elegance. The patient who was first treated with the experimental therapy, Jan Janisch-Hanzilik at age 49, had given up her active nursing career because her MS was reducing her mobility and she feared being unable to care for the grandchildren she loves. She called the clinic every other month until she was eligible to enroll. The treatment worked: her MS stopped advancing. Hundreds of clinical trials are now following her example, testing CAR T for lupus, Graves’ disease, vasculitis, and a growing list of other conditions. The treatment is not without risk — dangerous inflammation is a known acute side effect, and long-term safety data is thin — but no immunotherapy this effective has ever existed for most autoimmune patients before.

Google Co-Scientist and FutureHouse Robin: AI That Helps Scientists Think

On the AI-for-science desktop, two papers published simultaneously in Nature in May 2026 delivered what might be the most quietly inflectional moment for scientific AI since AlphaFold. Google’s Co-Scientist and the nonprofit FutureHouse’s Robin are both agentic AI science assistants, but they represent two different viewpoints on what "help" looks like.

Co-Scientist is designed as what Google calls a "scientist in the loop" tool: an AI that generates scientific hypotheses, and a human researcher periodically intervenes to direct its results. FutureHouse’s Robin goes a step further: it has been trained not just to generate hypotheses but to evaluate biological data from certain classes of experiment and return a reasoned assessment. For both systems, the guiding problem is one of scale. The number of peer-reviewed papers published each year has exploded past two million. Key cross-disciplinary connections — a signaling molecule active in eye development that’s also relevant to kidney research — are routinely missed by human researchers who simply cannot read every paper in every field. An AI that runs literature searches in the background and surfaces those connections is not a replacement for a scientist: it is a reading assistant so powerful that it changes what reviewing literature as a job actually entails.

The broader implication is that two-decade-old debates about AI replacing scientists have been largely superseded by what is actually happening: AI is infiltrating scientific workflow at the seams without fanfare, absorbing a specific category of cognitive-toil work that was already burning researchers out. Drug retargeting, hypothesis ideation, literature synthesis… these are jobs for which the current model of AI is specifically well-suited. The publishing of both systems simultaneously in Nature is Google’s and FutureHouse’s announcement that they’re no longer asking the question "can this work” — they’re asking "should we distribute it freely, or commercialize it through pharma partnerships.”

The Bits We’re Too Tired to Talk About Properly

Deepfake Accountability Starts (Thirteen Years Too Late)

Two people were criminally charged and unsealed in Brooklyn in late May 2026 under the Take It Down Act, which passed a year prior making it a criminal offense to post non-consensual intimate AI-generated deepfakes. The timing — exactly one year after the law’s prohibitions formally took effect, but long after the damage to victims’ lives had been done — is itself a commentary on the regulatory lag problem. More politically notable, the NTSB had to shut down a public accident docket after releasing a restoration image from a 2025 UPS airline crash: that image contained enough image-recognition exploitable signal that researchers could theoretically reconstruct audio from the cockpit voice recorder from a single image. The NTSB could not legally release such audio, so it had to pull the docket mid-stream. The incident is a very literal, very current argument for why "image recognition and computational methods" need regulatory guardrails in addition to the human decisions of regulators. No verdicts yet.

The Robotaxi and the Legislature

A bipartisan House bill proposes a $130 annual fee on EV owners because they don’t pay the gas taxes that fund US road maintenance. It’s not techno-optimism, it’s fiscal math catching up with a decade of electric transition. But it’s also structurally related: the vehicle platforms that will eventually run hands-free autonomous driving systems like Wayve’s are almost universally electric. The regulatory frameworks for EV taxation and the regulatory frameworks for supervised autonomous vehicles are being drafted in the same legislative calendar. That overlap��isusually managed in separate hearings by people who don’t talk to each other. Watch this space.

Looking Forward

May 2026 is not an outlier; it describes a new normal. The three axes of progress — AI moving from chat to systems, self-driving cars moving from demonstrations to OEM partnerships and regulated manufacturing, and biotech moving from concept to clinical application — are converging toward a world where the technologies of 2018’s science fiction require no imagination to understand, because you have now experienced them in one form or another in the past six months. The question measuring the next twelve months is not whether this progress will continue — it will — but whether governance, labor systems, and cultural frameworks can absorb it fast enough that it feels like progress rather than upheaval.

The chicken hatching from May’s most unexpectedly precise biotech egg, the passenger riding a Wayve-augmented FSD-equipped Stellantis vehicle on the highway, or the molecular biologist checking whether her journal-surfacing AI found a connection it couldn— they are all arriving at roughly the same rate, and the pace is not slowing.