Tesla, Microsoft and Anthropic all hit the same wall this week: the AI works, its infrastructure doesn't

By Zendoric · September 8, 2026.

Last Thursday Tesla began offering rides in its Cybercab, a two-seat robotaxi with no steering wheel or pedals, in limited areas of Austin. Less than 24 hours later, the U.S. federal road safety regulator (NHTSA) opened an investigation, as reported by TechCrunch and Reuters. That same week, Microsoft gave a name to its Windows for running AI locally and Anthropic acknowledged, through its own job postings, that it is building its own payments system. Three different stories, one and the same diagnosis: AI already works; the infrastructure around it does not.

The mechanics of the Tesla case matter. Federal vehicle safety standards still require manual controls, brake pedal included. Tesla self-certified the Cybercab as compliant —the usual procedure among manufacturers— but the NHTSA wants to review 'the process and technical data' on which that self-certification relied, and in particular whether Tesla decided that certain standards simply do not apply to a car without a steering wheel. The NHTSA's own administrator summed it up this way: 'we need to make sure all of our laws are being followed.' The Department of Transportation has already proposed eliminating the manual-controls requirement for vehicles designed to drive themselves, but as long as that rule does not change, the current one remains in force.

There is a precedent and it took four years: Amazon's Zoox self-certified its steering-wheel-free robotaxi in 2022, went through a special order and an NHTSA audit, obtained a demonstration exemption in 2025 and did not obtain final approval to charge fares until July of this year, with a cap of 2,500 vehicles a year for two years. It now charges fares in Las Vegas. The operational scoreboard, according to Texas records cited by Reuters: Tesla has 420 autonomous vehicles registered in the state, 45 of them Cybercabs; Waymo has 988. Tesla, moreover, has no permit in California to operate a robotaxi or to drive without a safety driver, and the July 2025 promise to have robotaxis available to half the U.S. population before the end of last year was not met.

None of this has slowed Tesla's valuation, $1.4 trillion according to Reuters, far above any other carmaker, sustained in good part by the autonomy narrative —the robotaxi on one side, the Optimus humanoid robots on the other— plus Elon Musk's bets on SpaceX and data centers. Waymo, with a valuation of $110 billion and a clear operational advantage —lidar and radar that work in bad weather, and the demand funnel from Uber and Lyft— is ahead on almost every number. China goes even further: Baidu has already surpassed 14 million trips and is expanding into Europe, Asia and the Middle East.

The second piece is less epic but just as revealing. Microsoft this week gave a name to its developer-oriented Windows, Project Zenith, with a clear entry requirement: unified memory —RAM and GPU memory shared on the same chip— of 64 gigabytes or more, according to The Verge. The first device is an AMD mini PC with its Ryzen AI Halo chips, unveiled at the IFA trade show; more machines with other silicon makers will arrive 'in the coming months.' Microsoft's promise is that developers will be able to run models of more than 30 billion parameters locally 'without a meter,' reducing dependence on cloud tokens billed by usage.

There is a convergence worth underlining: Nvidia's RTX Spark N1X chip arrives in October with up to 128 gigabytes of unified memory, and Apple's new Mac mini with the M6 chip starts at $899 with unified memory out of the box, a design Apple has used since the origin of its own silicon. Microsoft, Nvidia and Apple are drawing the same line by different routes: 64 gigabytes and up is the new floor for a computer that wants to run AI without depending on the cloud.

The third story is the one that says the most about the AI business. Anthropic is reinforcing its billing, fraud detection and finance infrastructure with in-house staff, according to its own job postings reported by The Information. One of them says it without beating around the bush: the 'business is growing faster than the processes and systems that support it.' The company's annualized revenue multiplied by nearly five to just under $45 billion in the five months through May, according to the same report. The problem is not selling the product, it is getting paid for it: complex token-based pricing (the token being the smallest unit of text a model processes, and the basis of its pricing) is far harder to bill and to protect against fraud than a fixed monthly subscription.

The openings spell it out: a role to rebuild order management and billing from top to bottom; a senior billing engineer with, literally, 'opinions about where [external platforms] fall short'; a fraud team to detect subscription and free-credit abuse in real time; and a treasury role to build 'production-grade financial applications that no vendor has built for us yet.' Anthropic notes that Stripe 'has been a solid partner for years' and that it continues to work with them across much of the business: this is not a breakup, it is a company beginning to build in parallel what it used to delegate entirely.

It is not an isolated case. As sector context, OpenAI added Adyen as a second payment processor last month and moved card data storage to an intermediary so it could work with several providers at once; Shopify brought in PayPal two years ago and has been accumulating state money transmission licenses. Stripe, for its part, has responded by moving up the value chain: it paid $7.5 billion for OpenRouter (a platform that charges those who buy access to AI models, not those who sell them) and $1 billion for Metronome, which specializes in usage-based billing; Adyen bought Orb for the same reason.

The three stories share the same structure and, for Zendoric, tell the same underlying news: AI capability keeps arriving before the infrastructure that supports it. The road is not ready for a car without a steering wheel; the consumer computer is not ready to run an AI agent locally without a memory upgrade; and payment networks are not ready for a company that charges per token at a scale of billions of operations. This is not an anomaly of this industry: the PC arrived before the software that justified it, broadband arrived after the internet had already gone mainstream, the app economy arrived years after the first smartphone. The infrastructure layer, the one that does not make the headlines, is always built after the product on top proves it is worth it.

In the short term this is real friction, not a technical detail. That the NHTSA opens an investigation within hours of Tesla starting to offer rides in a car with no manual controls is exactly what a regulator should do: hit the brakes before safety gets proven on the fly with passengers on board. The Zoox case itself —four years from self-certification to final approval to charge fares— confirms that unsupervised autonomous driving advances in years, not quarters, something we have been pointing out for some time; Tesla's scoreboard in Texas against Waymo, and the absence of a permit in California, indicate that Tesla's stock market narrative is running ahead of its actual deployment.

The shift by Microsoft, Nvidia and Apple toward unified memory as an entry standard has a more optimistic reading, and it connects with something we argue at Zendoric: running models with tens of billions of parameters on a desktop computer, with no cloud token meter, is a democratizing force. It cuts costs, returns control of data to the user or the company, and fits with the rise of open models that already compete head to head with closed ones. The entry price —machines with 64 gigabytes or more of memory, in the middle of a RAM shortage— remains high today, but the fact that three companies are converging in the same direction suggests that floor will come down over the years.

The Anthropic case is perhaps the most eloquent about the real state of the AI business. A company that multiplies its annualized revenue by nearly five in five months to just under $45 billion does not have a demand problem: it has a financial plumbing problem, because nobody —not even a provider as established as Stripe— has yet built payment systems designed for micropayment billing per token at massive scale. The newsletter that reports this story also points to a possible IPO at a valuation above $2 trillion in the coming weeks; that is a figure circulating as a market expectation, not a fact confirmed by Anthropic, and it should be treated with that caution.

This is where Zendoric's underlying thesis fits: the internet was not born with micropayments either —HTTP status code 402, 'payment required,' was reserved in the nineties and was never actually built— and even so digital commerce found its own patches, from PayPal to Venmo, over thirty years. Now the industry is repeating the exercise for an economy of tokens billed in fractions of a cent, with an even larger horizon ahead: AI agents paying each other, at machine speed, with no human intervention. Building that plumbing —regulatory, hardware and financial— is the least glamorous work of this AI wave, but it is the work that determines whether the abundance the technology promises arrives in years or in decades. The capability is already here; what is missing are the pipes.