Sovereign AI, same host: Thoughts on the race of physical AI

Physical AI developers determined not to rely on NVIDIA have scored some victories lately. OpenAI provided its own chip. Anthropic has built an internal silicon team with salaries that make FAANG recruiters wince.

Additionally, Advanced Micro Devices acquired World Labs for $8.2 billion, Qualcomm partnered with Liquid AI on embedded agents, and Skild AI launched the S1 foundation model. Additionally, Google DeepMind integrated Gemini Robotics 2 into Apptronik’s Apollo 2 humanoid.

And then, almost politely, NVIDIA said it planned to spend $12.9 billion – the largest acquisition in the company’s history – buying Hugging Face, the place where the planet’s developers go to choose which model to run. Cute war. NVIDIA just bought the battlefield.

Chip hunting focuses on the wrong layer

Several headlines read “Labs race to get rid of NVIDIA.” Lovely shot. Here’s the real scoreboard: While everyone else was designing chips, NVIDIA designed an off-ramp that leads right back there, regardless of which chip wins. This is not defense. It’s a company that has already read the ending and is waiting, patiently, for the rest of the industry to catch up.

The chip numbers are really impressive, to be honest. OpenAI spent nine months with Broadcom on Jalapeño. This is reportedly the fastest tape-out in the history of advanced semiconductors, with delivery later this year.

Anthropic followed with its own silicon team, posting hardware roles worth up to $485,000, piled on a multi-gigawatt TPU (Tensor Processing Unit) deal it already had in place with Google and Broadcom. Google has had TPUs for a decade.

AWS has Trainium. Meta has MTIA. Everyone is rushing to stop writing a check to Jensen Huang.

None of these very expensive chips reach the level where the other 200,000 companies that will never build their own silicon actually go to find, tune, and ship a model. That level, the true starting point for AI adoption in the rest of the world, has just been quietly annexed by the company everyone claims to have staged a jailbreak from. NVIDIA didn’t bat an eyelid; he bought the prison.

‘Sovereign’ is a clue to NVIDIA’s strategy

Note the word that NVIDIA has continued to use in every earnings call so far this year: “sovereign.” Not developers first. Not businesses first. Sovereigns: Nation-states spend public money on “AI independence” so they are not permanently kneeling before a handful of US labs.

No government building sovereign AI wants to bet its national computation on OpenAI’s chip, Anthropic’s silicon, or any single lab that might one day become a geopolitical liability. Then NVIDIA stepped in with the message: “Don’t worry, we’re neutral. Choose any model, any framework, any cloud you want” – a rousing speech about freedom of choice, delivered by the company that just bought the store where all the choices reside.

This is the real chess move. OpenAI and Anthropic are optimizing their cost of goods. NVIDIA is optimizing to stay below the cost of everyone’s goods, forever, regardless of who wins the battle it’s graciously letting other people have. No one at Nvidia will ever put this on an earnings call. It is not needed. The acquisition forms their argument.

Here’s the part no one wants to say out loud: Even a great custom chip takes years to match the volume, throughput, and software maturity of a platform that NVIDIA spent two decades pouring into CUDA.

Jalapeño’s nine-month tape-out is true engineering flexibility. But by all accounts, it’s still a narrow inference chip built for OpenAI workloads, not a platform that someone else is building a business on. Anthropic’s silicon team is, charitably, a hiring page with a mission statement, nestled within a “multi-chip strategy” that still runs today primarily on NVIDIA and Google hardware.

Meanwhile, NVIDIA just launched a distribution channel with Hugging Face that receives more daily developer traffic than both labs’ entire platforms combined.

The sovereign implications become more acute for robotics

The implication becomes more acute in physical AI. A robot can’t wait for a distant API to deliberate while it moves through a warehouse, operates sensitive equipment, or makes a safety-critical decision in a factory.

It needs local inference, real-time control, sensor data that stays within the operation, and a delivery stack that doesn’t collapse due to a hiccup in a cloud region or a vendor changing its terms. This is precisely where “sovereign” stops being a slogan in procurement and becomes an operational requirement.

But the same problem as the landlord follows the robot to the factory. The company that owns the training stack, simulation environment, edge runtime, development tools, and model marketplace isn’t just selling silicon, it’s defining the terms of how physical intelligence is built and deployed.

Physical AI makes hardware choice more important, not less. It also makes true portability between models, compute and deployment environments much more valuable.

None of this changes much for the companies that ship products to customers—if anything, it makes the point clearer. Airrived last month launched its sovereign AI platform, positioning itself as an agent operating system with a mission to democratize AI and turn every business user into an AI expert.

This is the other, more literal definition of “sovereign”: not a government hedging its national cyber bet on a chip supplier, but a business user who doesn’t need a data center, a silicon roadmap, or a cutting-edge engineer to own and operate its own AI.

NVIDIA’s sovereignty version comes with an owner attached. The application level version doesn’t care what logo is silk-screened onto the underlying silicon. He cares if the inference is fast, cheap, and boringly reliable, and he’s happy to let adjacent $2 trillion companies fight over who gets to foot the infrastructure bill. This is the hidden benefit of building a lean approach on top of the war instead of funding a part in it: the fight against the chips is enormously expensive to win and completely free to ignore.

Where silicon is actually cooked, this isn’t shaping up to be the underdog story the press releases make it out to be. OpenAI and Anthropic are spending billions to buy independence from NVIDIA hardware. NVIDIA just spent $12.9 billion to buy insurance against the loss of either, owning exactly where their models need to appear regardless of the chip running them.

Call it war if you want. From here, it looks like NVIDIA is collecting rent on both sides of the battlefield, smiling and asking if anyone else would like to diversify.

About the author

Anurag Gurtu, co-founder and CEO of Airrived, is an entrepreneur and product leader with over 20 years of experience in cybersecurity and artificial intelligence. He has created three generative AI companies, including one that has achieved unicorn status.

Gurtu is known for transforming advanced artificial intelligence into scalable, real-world platforms and has previously contributed to successful exits, including acquisitions by Splunk, Tripwire, Sun Microsystems, and FireEye, a public company. He holds an MS from the University of Southern California and CISSP certification.