US AI Pre-Deployment Testing: Google, Microsoft & xAI

Samantha: Welcome to The State of Tech, The European Edition, Wednesday May six, 2026. I'm Samantha Lawrence.

Bob: And I'm Bob Russell. Today: Google, Microsoft and xAI hand the US government early access to their frontier AI models; European lawmakers demand urgent action after the Mythos hacking revelations; a Singapore chip fab pivots to AI interposers; Google patents that aim to keep AI data centres from going dark; NAND flash memory's surprise comeback in the AI era; and a hybrid robot from the UAE built for the rough stuff. Let's start with that landmark Washington deal.

Google, Microsoft and xAI agree to let the US government test their frontier AI models before release.

Samantha: This is a genuine shift in how the most powerful AI systems get vetted. Google, Microsoft and Elon Musk's xAI have agreed to give the US government pre-deployment access to their top models for safety evaluations. The deal was announced by the National Institute of Standards and Technology, through its Center for AI Standards and Innovation, known as CAISI.

Bob: And the trigger is no secret. It's the capability that Anthropic's Mythos system showed in cybersecurity, autonomous hacking, vulnerability discovery at a scale that genuinely rattled people in Washington. CAISI now wants to look under the hood before these models hit the market, not after.

Samantha: Notably, Anthropic isn't on this list of three. Whether that's because Mythos is already being assessed separately, or because the relationship is more complicated, nobody at NIST is saying out loud. But the three names that did sign on cover an enormous slice of the frontier.

Bob: The scope of testing matters too. We're talking national security risks, public safety implications, misuse in cyber operations, potentially bio risks. It's voluntary in form, but in practice if you want to operate in the US market with a frontier model, this is becoming the price of entry.

Samantha: And it reframes the conversation around AI governance. For two years the debate has been: do we regulate before harms, or after? Pre-deployment testing is a clear answer, before. That's a meaningful philosophical shift for an American administration that has generally preferred a lighter touch than Brussels.

Bob: Which brings us straight to the European angle.

Samantha: The AI Act already foresees evaluations for general-purpose models with systemic risk, but the operational machinery isn't fully there yet. If American agencies are getting weights and red-team access months before launch, European regulators won't want to be the last to see the thing.

Bob: There's also a sovereignty question. European frontier developers, Mistral, Aleph Alpha, Silo, will face the question: do you submit to American testing if you want US distribution, and what does that mean for the IP? Brussels will want a reciprocal arrangement, not a one-way mirror.

Samantha: And European critical infrastructure operators, banks, energy grids, hospitals, will want to know what the testing actually found. Right now, those reports stay inside the US government. That's the next fight.

European MEPs demand urgent cybersecurity action after the Mythos AI hacking revelations.

Bob: Staying with Mythos, because the European Parliament is now formally alarmed. Dozens of MEPs have written to the Commission demanding stronger cybersecurity measures in direct response to what Anthropic's Mythos has demonstrated, and to reports that Mythos itself was subject to unauthorised access.

Samantha: That second part is the kicker. It's one thing to worry about a powerful offensive AI in the right hands. It's quite another to learn the model may have leaked from the lab. The MEPs are saying, in effect, the threat has stopped being theoretical.

Bob: Their specific complaint is that Europe lacks institutional access to Project Glasswing, which is Anthropic's defensive research strand around Mythos. Washington has a relationship there. Brussels does not. So European cyber defenders are reading about capabilities they cannot themselves test against.

Samantha: And the MEPs name the obvious targets, public services, hospitals, water utilities, energy operators, transport. The asymmetry right now is uncomfortable. Attackers can rent or steal frontier capability. Defenders are still budgeting for last year's threat model.

Bob: There's a policy ask underneath this letter. They want emergency funding, they want ENISA scaled up, and they want the Commission to negotiate access to frontier defensive AI for member-state CSIRTs. That's a significant expansion of what European cybersecurity policy normally covers.

Samantha: The political timing is sharp. The Commission was already preparing a cyber resilience update. This letter forces them to address frontier AI risk explicitly, not bury it in a footnote.

Bob: And it raises a harder question for European AI strategy. If Brussels wants its own defensive Mythos equivalent, it needs the compute, the data, and the talent. None of that is solved by a letter.

Samantha: Quick interruption. If you listen to The State of Tech regularly, hit that like button and subscribe, that way you'll never miss an episode. Okay, moving on.

Vanguard repurposes its new Singapore fab to make interposers for AI chips, with TSMC tech.

Bob: Vanguard International Semiconductor, VIS, is reconfiguring its brand-new twelve-inch fab in Singapore. Part of that capacity is being redirected to manufacture interposers, the silicon layer that sits between an advanced AI chip and its package, routing thousands of connections between compute dies and high-bandwidth memory.

Samantha: And they're doing it by licensing the relevant technology from TSMC. That's notable. TSMC's CoWoS packaging has been the bottleneck for AI accelerators for two years. Anything that adds interposer capacity in the supply chain matters.

Bob: The trade-off is visible in the numbers. Total installed capacity at the fab has been scaled down to forty-four thousand wafers a month, because interposer production is more complex and consumes more floor space per wafer. Volume production is targeted for the first quarter of next year.

Samantha: The fab will also produce mixed-signal, power management and analog chips. Which tells you something about the economics, AI interposers are the headline, but you still need the unglamorous parts that go into every car, every phone, every appliance.

Bob: Singapore as a location is a deliberate hedge as well. It diversifies advanced packaging away from Taiwan, which has been the geopolitical anxiety underlying the entire AI hardware industry.

Samantha: European hyperscaler projects, the EuroHPC AI factories, the sovereign cloud builds, all depend on getting their hands on accelerators that themselves depend on interposers. A bottleneck in Taiwan ripples directly into delivery dates in Frankfurt and Paris.

Bob: It also reinforces the case for the European Chips Act second phase. Europe has wafer fabs on the way, but advanced packaging is the gap. If Singapore can license TSMC tech, in principle so can a European partner, but only if the funding is there.

A Google engineer wins four US patents for keeping AI data centres powered through anything.

Samantha: Krishnanjan Gubba Ravikumar, an engineering director at Google, has been granted four US patents this year focused on power resilience for AI data centres. The patents cover distributed power transfer, dynamic anomaly protection, energy control systems, and uninterruptible power supply timing.

Bob: This sounds dry until you remember that an AI training cluster that loses power mid-run can lose days of progress and millions of dollars. Power density in AI data centres has climbed past one hundred kilowatts per rack in some deployments. The old UPS playbook doesn't really hold at that scale.

Samantha: The dynamic anomaly protection piece is interesting. Instead of waiting for a fault and tripping a breaker, the system tries to detect the precursor signature, voltage drift, harmonic distortion, and isolate the affected zone before the GPUs notice anything.

Bob: And the distributed power transfer patent essentially treats the data centre as a small grid that can route around problems internally. If one substation feed wobbles, load shifts to another path within milliseconds. TPUs don't even hiccup.

Samantha: There's a broader story here about who actually controls the standards for AI infrastructure. Patents like these become the de-facto reference design, even for operators who don't licence them, because the architecture leaks into industry norms.

Bob: Ireland, the Netherlands and parts of Germany have already imposed grid connection limits on new data centres. If you can't get more power, you have to use what you have more reliably. These Google techniques fit that European reality.

Samantha: And there's a sustainability angle. Smarter power management means less spinning reserve, less diesel backup running idle, lower overall energy waste. That helps with the EU energy efficiency directive, which now applies directly to data centres above a certain size.

NAND flash memory, written off a year ago, is now the unsung backbone of the AI era.

Bob: NAND flash, the memory that goes into SSDs, has had a brutal couple of years. Oversupply, weak phone sales, prices through the floor. Now it's roaring back, because AI training and inference need eye-watering amounts of fast storage.

Samantha: HBM, the high-bandwidth memory bolted to GPUs, gets all the attention. But the data those models train on has to live somewhere, and increasingly that somewhere is enterprise NAND. Datasets in the petabyte range are now routine for frontier training runs.

Bob: Samsung and SK Hynix are the obvious winners, and analysts expect a meaningful profitability swing as the industry moves to tenth-generation NAND next year. That's higher density, lower cost per bit, better power efficiency, exactly what AI buyers want.

Samantha: There's also a shift in how NAND is bought. Hyperscalers are now signing multi-year supply agreements with floor pricing, similar to what they already do for HBM. That changes the cyclicality of the memory industry, which has historically been a boom-bust horror show.

Bob: From ugly duckling to essential component, as the analysts put it. Which feels right. Storage was always going to matter, the question was just when the market would notice.

Samantha: If the hyperscalers lock in long-term NAND supply, smaller European buyers risk paying spot prices that climb steadily through the second half of this year. That hits sovereign cloud projects and university clusters hardest.

Bob: It's also a reminder that Europe has almost no domestic memory manufacturing. We talk about chip sovereignty, but memory has been quietly written out of that conversation. This rebound makes that gap more visible.

A UAE robotics firm unveils a hybrid all-terrain robot built for the messy real world.

Samantha: To finish on something a bit more fun. Micropolis Robotics, based in the UAE, has unveiled its M1.5 hybrid robot at the Make it in the Emirates 2026 event in Abu Dhabi. It's an autonomous mobile platform built for environments that polite indoor robots simply can't handle.

Bob: Think gravel, sand, slopes, long patrol routes in heat that would melt a delivery bot. The M1.5 has improved off-road mobility, longer endurance, and what the company calls operational resilience, which is marketing speak for, it doesn't break when you actually use it.

Samantha: It's pitched at security, surveillance, and smart infrastructure. Patrolling industrial sites, ports, oil and gas facilities, large solar farms. The kind of jobs where a human guard in a vehicle is expensive, bored, and often missing the actual incident.

Bob: The hybrid bit refers to the drivetrain, combining electric operation for quiet, efficient running with a backup that extends range well beyond what a pure battery platform manages. For a security robot that needs to work all night, that matters.

Samantha: And there's a quiet trend underneath this. The Gulf is becoming a serious robotics market, partly because there's capital, partly because there's appetite, and partly because the climate is genuinely hostile to humans doing outdoor work for hours at a time. Robots there are not a gimmick.

Bob: It also pushes the category forward. Most autonomous mobile robots you see today are warehouse creatures, smooth floors, predictable layouts. Building one that can handle a construction site or a desert perimeter is a different engineering problem entirely.

Samantha: European ports, Rotterdam, Antwerp, Hamburg, are already piloting outdoor autonomous patrol units. A platform like the M1.5 fits that brief. It also raises the usual questions about surveillance, data handling and cross-border deployment of robots that carry cameras and sensors everywhere they go.

Bob: Which is a properly European thing to worry about. Even our robots get GDPR.

Samantha: Today we covered: the US pre-deployment AI testing deal with Google, Microsoft and xAI; European MEPs sounding the alarm after Mythos; Vanguard's AI interposer pivot in Singapore; Google's data centre power patents; NAND flash flipping from laggard to AI essential; and Micropolis unveiling its rugged M1.5 robot.

Bob: Want to know more or react? Visit stateoftech.eu or email us at info@doorzetters.net.

Bob: State of Tech, the tech world in 15 minutes.