[Weekly Roundup] Nvidia is now underwriting its own customers
Morning. Five stories this week, and four of them are really the same story wearing different clothes: the money and the machinery behind AI are getting very large, very fast, and the arrangements holding it together are starting to look unusual.
Computing
Nvidia is reportedly guaranteeing 100 billion dollars of OpenAI's credit
The chip seller is helping underwrite the chip buyer. Nvidia is said to be close to backing roughly 100 billion dollars in credit for OpenAI's next data centre build. The money leaves through the front door and comes back through the loading bay as GPU and networking orders.
Why it matters: Vendor financing at this scale makes demand look bigger than it is, because part of that demand was created by the supplier. Telecoms equipment makers ran this exact play during the fibre boom. It worked brilliantly until it stopped.
Chips
Google bought its way deeper into custom AI silicon
A deal reported at 12.2 billion dollars could hand Google a stake in Marvell. Marvell does the unglamorous work that turns an accelerator design into a manufacturable system: interconnect, memory controllers, optical links, packaging. Google has run TPUs internally for a decade and is now selling that capacity to anyone who wants it.
Why it matters: Nobody displaces Nvidia at the frontier this year. But high volume, well understood inference is exactly where custom chips earn their keep, and that is the boring revenue underneath the whole market.
AI
Apple trained a separate AI model for China with Alibaba
One phone, two AI stacks, decided by where you bought it. Chinese rules require generative models offered publicly to be registered and filtered, so Apple could not simply route requests to servers in California. Alibaba supplied the model expertise, the infrastructure and the licences.
Why it matters: Ask the same question of the same handset in Shanghai and in Seattle and you may get different answers by design, not by accident. The EU AI Act, China's registration regime and a growing pile of national data rules are all pushing towards regional variants of what used to be one global product.
Transport
Uber and Pony.ai are putting 2,000 robotaxis on European roads
Driverless taxis are leaving the wide grids they grew up on. Phoenix and Shenzhen are generous to autonomous vehicles. A medieval street in Bologna, a Dublin junction with temporary lights and a cyclist coming the wrong way down a bus lane are a different proposition. There is also no single approval to win, because operating rules, insurance and liability are set country by country.
Why it matters: Two thousand vehicles across a continent is a rounding error next to the number of drivers working today. The number that matters is cost per mile once the safety driver is gone, and how fast regulators allow that. Also worth watching: Pony.ai is a Chinese firm, and a fleet constantly mapping European cities will attract exactly the data questions you would expect.
Security
Officials are warning that hackers are hitting water and energy controllers
The kit running water treatment and power distribution was never designed to be online. Programmable logic controllers were built to sit in a cabinet for twenty years, and plenty of them ship with default credentials and protocols that assume everything on the wire is friendly. Then somebody connects the plant for remote monitoring.
Why it matters: Most of you do not run a substation, but these intrusions usually start with a stolen login belonging to a contractor or engineer. That part is defensible. A phishing resistant hardware key makes a stolen password worthless, and setup takes about ten minutes. Check the YubiKey 5 NFC on Amazon
One to watch next week
Amazon is preparing the largest expansion of its Prime Air drone network so far, aiming at close to 500 US cities and towns by the end of the year. That is roughly six times its current footprint. The interesting part is not the drones, it is the aviation regulators: the US approach to beyond visual line of sight flight has moved faster than most of Europe and Asia, and if this rollout holds up, expect regulators in the UK, EU and Australia to face loud questions about why their own approvals are slower. Announcements from Amazon typically land in the US morning, which is afternoon UK time and late evening across East Asia.
Forwarded this? Get your own free briefing at futuretechnologyhq.com/newsletter
Nath, Future Technology
Some links in this newsletter may be affiliate links. We only recommend products we genuinely think are worth your time.