- HEADLINE & INTRO
Uber’s latest cloud decision is less about ride‑hailing and more about who will own the foundations of the AI era. By deepening its reliance on Amazon Web Services and trialing Amazon’s in‑house Trainium3 AI chips, Uber is quietly endorsing a future where cloud providers win not just on data centers and software, but on silicon. This move exposes how fragile even “mega” multi‑cloud deals can be, and why custom chips are becoming the sharpest weapon in the fight against Nvidia – and against rival clouds.
In this piece we’ll unpack what Uber actually signed, why it matters far beyond one contract, and what it signals for AI costs, competition, and European users.
- THE NEWS IN BRIEF
According to TechCrunch, Amazon announced that Uber is expanding its use of AWS infrastructure to power more of its ride‑sharing workloads on Amazon‑designed chips.
Concretely, Uber will increase adoption of AWS’s Graviton processors – ARM‑based general‑purpose CPUs known for lower power consumption and better price‑performance – and launch a new pilot on Trainium3, Amazon’s latest AI accelerator positioned as an alternative to Nvidia GPUs for training large models.
This is notable because in 2023 Uber signed high‑profile, multi‑year cloud deals with both Oracle Cloud Infrastructure and Google Cloud, with the stated goal of moving most of its on‑premise infrastructure to those two providers. Uber has publicly highlighted Oracle’s ARM instances based on Ampere chips as part of that migration. Since then, Ampere has been acquired by SoftBank, while Oracle exited its stake.
Now AWS is effectively inserting itself into that picture by winning more of Uber’s workloads specifically on the strength of its proprietary chips.
- WHY THIS MATTERS
The obvious winner is Amazon. It doesn’t just gain more Uber workloads; it gains a high‑visibility reference customer validating its silicon strategy. For years, Graviton and Trainium were seen as interesting side bets. When a hyperscale, latency‑sensitive platform like Uber leans into them, it tells every other large customer: these are production‑grade, not experiments.
Uber also stands to benefit. Ride‑hailing, food delivery and logistics are low‑margin, volume businesses. AI is now embedded everywhere in Uber’s stack: ETA prediction, pricing, driver matching, fraud detection, customer support. If Graviton and Trainium3 deliver the promised performance per dollar, Uber gets cheaper inference and training – and, crucially, leverage in negotiations with Oracle, Google and Nvidia.
The potential losers are more subtle. Nvidia won’t feel immediate pain from one Trainium3 pilot, but the direction of travel is clear: major buyers are actively cultivating non‑Nvidia options to escape GPU pricing and supply constraints. Oracle looks worse: it sold its stake in Ampere and publicly de‑emphasised in‑house chip design, just as one of its marquee customers is being courted precisely on the strength of a rival’s home‑grown chips.
Strategically, the deal underlines a bigger shift: in cloud, the new lock‑in is not just APIs or managed databases – it’s the chip stack. If your AI workloads are tuned to Trainium or Graviton, re‑platforming to another cloud becomes harder, or at least more expensive. AWS is building a moat from the transistor upward.
- THE BIGGER PICTURE
Uber’s move fits into a broader industry pattern: hyperscalers are racing to own their silicon destinies.
Google has had its TPU line for years. Microsoft is rolling out its own AI accelerator (Maia) and ARM CPU (Cobalt). Meta has its MTIA chips for inference. Amazon, arguably the most vertically integrated of them all, now has a multi‑billion‑dollar business around Graviton and Trainium, as its CEO highlighted at the end of 2025.
The common driver is Nvidia dependency. The AI boom turned Nvidia into a single point of failure for the entire sector: supply bottlenecks, skyrocketing costs, and a roadmap that puts Nvidia – not the cloud providers – in the strategic driver’s seat. Custom chips are the counter‑move: they don’t need to beat Nvidia on raw performance, only on performance per watt and per euro for specific workloads.
Historically, this mirrors how the public cloud itself emerged. In the 2000s, owning data centers was a key differentiator; then it became table stakes. Today, racks of generic x86 servers are a commodity. The new premium layer is accelerators and tightly coupled software stacks.
The Ampere saga, which TechCrunch recounts, illustrates the volatility of this space. Oracle invested heavily in an ARM vendor to differentiate its cloud, only to sell its stake and declare in‑house chip design a distraction, just as custom silicon becomes central again. SoftBank picks up Ampere, looking to assemble its own AI infrastructure empire. Uber, caught in the middle, is now effectively hedging its bets across three clouds and multiple chip line‑ups.
In other words: the AI infrastructure market is fragmenting at the silicon level while consolidating around a handful of hyperscalers. That tension will define the next decade.
- THE EUROPEAN / REGIONAL ANGLE
For European companies and public institutions, Uber’s decision is a reminder of a hard reality: most cutting‑edge AI infrastructure will, for the foreseeable future, live inside US‑based hyperscalers and their proprietary chips.
From a purely economic standpoint, more chip competition is good news. If Trainium3 and Graviton force Nvidia to moderate prices, European AI startups and research labs gain. Many of them already run on AWS credits or similar programmes; lower training and inference costs lengthen their runway.
But there’s a sovereignty trade‑off. Custom chips deepen vertical integration: hardware, runtime, model hosting and data all under one US corporation. That sits uneasily next to EU policy goals embedded in the Data Act, the Digital Markets Act and the AI Act, which all push for portability, interoperability and reduced gatekeeper power.
EU‑based cloud providers – from OVHcloud to regional players in Germany, France, the Nordics and CEE – mostly rely on standard CPUs and off‑the‑shelf GPUs. They simply cannot yet match the economies of scale of AWS‑style custom silicon. Their counter‑narrative is openness (no proprietary accelerators), strong data localisation and closer alignment with EU regulation.
For European regulators, the Uber–AWS news is another data point suggesting that “chip‑driven lock‑in” could become as serious as software or data lock‑in. Expect more scrutiny of how easy it really is to move AI workloads off proprietary accelerators, and whether interoperability obligations should extend down to that layer.
- LOOKING AHEAD
What happens next depends on three questions.
First, does Trainium3 actually deliver? If Uber’s pilots show real‑world gains in cost and reliability versus Nvidia‑based setups, we should expect Uber to expand its AWS footprint further – and competitors in mobility, logistics and delivery to at least test the waters. If performance disappoints, the deal risks becoming more of a marketing story than an architectural shift.
Second, how far will "chip‑centric" multi‑cloud go? Uber is now effectively multi‑cloud at three levels: providers (AWS, Oracle, Google), architectures (x86, various ARM designs), and accelerators (Nvidia, Trainium, potentially others). That’s powerful for negotiation, but complex to operate. Over time, most large customers gravitate toward one “anchor” cloud plus one or two tactical partners. Uber’s choice of anchor – decided as much by chip roadmaps as by generic cloud services – will be telling.
Third, how will Nvidia and the rest respond? Nvidia is already pushing its own cloud‑adjacent offerings and tighter integration with partners like Oracle. If AWS proves that custom chips can peel away high‑profile AI workloads, expect Nvidia to double down on software ecosystems (CUDA, AI frameworks) and potentially more direct services to customers.
For readers, the signals to watch over the next 12–24 months are clear: the pace of Trainium3 adoption beyond AI natives like Anthropic, disclosures from large enterprises about chip mix in their clouds, and any regulatory moves in the EU or US targeting hardware‑level lock‑in.
- THE BOTTOM LINE
Uber deepening its alliance with AWS around Graviton and Trainium3 is not just another cloud contract; it’s a vote for vertically integrated AI stacks where the most strategic decisions happen in silicon. That strengthens Amazon, pressures Nvidia and Oracle, and narrows realistic choices for anyone building serious AI in Europe and beyond. The open question is whether regulators – and customers – are comfortable letting chips become the next great chokepoint of the cloud era.
