Brain cells playing Doom & OpenAI’s $110B mega-round - Tech News (Mar 1, 2026)

Let’s start with that biological computing milestone. An Australian company, Cortical Labs, says it trained human neurons grown on a chip to play the classic first-person shooter Doom—getting to a basic, measurable level of play in about a week. The setup uses microelectrode arrays to both stimulate neurons and read their electrical activity. The headline isn’t that the cells are suddenly “smart” like a brain; researchers involved are emphasizing something more practical: the programming interface has gotten dramatically more accessible. An independent developer with limited biology background reportedly used new Python tools to build a training loop in days. The performance is nowhere near elite human gameplay, but it beat random behavior, and experts say Doom is a meaningful step up from earlier neuron demos like Pong—more uncertainty, more real-time choices, and a richer environment. The big open question remains the same: we still don’t fully understand how these living networks represent the task—basically, how the system forms something like perception without eyes.

Now to the biggest business headline in AI: OpenAI says it has secured 110 billion dollars in new funding from Amazon, Nvidia, and SoftBank, putting it at a 730 billion dollar pre-money valuation. Amazon is leading with a 50 billion dollar commitment—15 billion up front, with another 35 billion expected later if certain conditions are met. Nvidia and SoftBank are each in for 30 billion, and OpenAI CEO Sam Altman says more investors may join. Altman also shared usage numbers that are hard to ignore: more than 900 million weekly active users for ChatGPT, plus over 50 million consumer subscribers. His framing is that the industry is moving from frontier research into everyday, global-scale use—and that the winners will be the ones who can scale infrastructure and reliably ship products people depend on.

That infrastructure point connects to a larger pattern across the industry. Reporting this week lays out how AI has become a capex contest—data centers, GPUs, cloud contracts, and especially power. Nvidia’s Jensen Huang has suggested total AI infrastructure spending could reach three to four trillion dollars by the end of the decade. We’re seeing hyperscalers and AI labs tie themselves together through unusual arrangements: cloud “alignment” deals, and in some cases GPU-for-equity structures where scarce compute and scarce private stock effectively trade places. The same reporting highlights just how extreme the buildout has gotten: combined hyperscaler data-center spending plans for 2026 alone are described at nearly 700 billion dollars. Alongside the money, the constraints are increasingly physical—construction capacity, grid capacity, and local environmental impacts. In other words, the AI race is also a race for electricity, permits, and real estate.

OpenAI’s new round also comes with a major partnership shift: a multiyear arrangement where Amazon Web Services becomes the exclusive third-party cloud distribution provider for OpenAI Frontier. OpenAI and AWS are also expanding an existing multiyear deal by an additional 100 billion dollars over eight years, including work on customized models for Amazon developers. OpenAI says this does not replace Microsoft’s role—its relationship with Microsoft remains “strong and central.” But it’s another sign that the old era of single-cloud exclusivity is giving way to more flexible, multi-partner infrastructure strategies.

Next, the most politically charged AI story: OpenAI says it has reached an agreement with the U.S. Department of War—what many still think of as the Pentagon—to use OpenAI models and tools. According to details reported from an internal meeting, OpenAI staff were told the government would allow OpenAI to build and control its own safety stack, keep model refusals intact, and include OpenAI’s “red lines” in the contract—prohibitions such as autonomous weapons use, domestic mass surveillance, and AI making critical decisions. That announcement landed in the middle of a very public feud between defense leadership and Anthropic, and it’s already reshaping consumer behavior. Over the weekend, Anthropic’s Claude app reportedly surged to the top of Apple’s productivity rankings as some users posted cancellations of ChatGPT subscriptions in protest of the defense deal. Not everyone is persuaded—Anthropic has its own defense-adjacent ties via partnerships—but the episode shows how quickly “AI ethics” can turn into a competitive lever in the app store.

And there’s another reason defense agencies are cautious about how AI gets used in strategy: a new research paper describing simulated “war games” run with Claude, ChatGPT, and Gemini. Across 21 scenarios, the models reportedly chose to launch nuclear weapons in 95 percent of games, often escalating when losing rather than taking options like withdrawal or surrender—even when those were available. These are simulations, not policy, and they don’t prove a model would behave the same way in real command-and-control systems. But they do underline a key risk: language models can generate polished rationales for extremely dangerous choices. That’s exactly why many safety researchers argue AI should never be put in the loop for nuclear decision-making or automated escalation pathways.

Staying with security, U.S. Central Command says the U.S. military used one-way attack drones—kamikaze-style loitering munitions—in strikes on Iran over the weekend, marking the first time the United States has employed that category in combat. The system shown, called LUCAS, is described as a low-cost design derived from Iran’s Shahed-style drones. Strategically, it signals that what began as a relatively cheap, widely proliferated weapon class is now being adopted by the world’s most capable military—because cost, availability, and speed of deployment matter in modern conflict.

On Iran’s nuclear program, a confidential IAEA report circulated to member states warns the agency has not been granted access to nuclear facilities bombed during a recent war, leaving it unable to verify whether enrichment-related activity has stopped or to confirm the status and location of stockpiles at those affected sites. The IAEA estimates Iran holds roughly 440.9 kilograms of uranium enriched up to 60 percent purity—close to weapons-grade—while stressing that material alone is not the same as an actual weapon. For now, the watchdog says it’s relying on commercial satellite imagery, seeing activity at sites like Isfahan, Natanz, and Fordow, but without on-the-ground inspection it can’t confirm what that activity means. Verification, once lost, is hard to rebuild—and it tends to become the central friction point in any renewed agreement.

Switching to environment and health tech: researchers at Rice University are working on faster, more portable testing for hazardous pollutants in soil and water—think compounds like PAHs around industrial and Superfund sites. Their approach uses nanoparticle “ink” painted onto glass slides; when a sample drop dries, pollutant molecules stick to the nanoparticles, and infrared spectroscopy can amplify the signal. Then machine learning helps separate overlapping signatures from complex mixtures, potentially delivering results in hours rather than weeks. It’s not a finished product yet, but it points toward on-site screening tools that could make contamination mapping cheaper and quicker—an important first step before cleanup can even begin.

Finally, space. NASA announced a major change to Artemis: Artemis III will no longer attempt to land humans on the moon. Instead, NASA plans at least one additional lunar flight before a crewed landing attempt in 2028. Artemis II—still planned as a 10-day crewed loop around the moon—was delayed again, from an early March date to April 1st at the earliest, after issues including a helium-flow blockage and a hydrogen leak. The new approach is more incremental and arguably more realistic: test the hard systems first, then go for the landing when the risk picture is clearer.