Amazon’s $200 Billion Grudge Match: From Nvidia Customer to Full-Stack Rival

Amazon’s $200 Billion Grudge Match: From Nvidia Customer to Full-Stack Rival

Amazon’s latest shareholder letter is less a reassuring memo and more a declaration of technological independence. Andy Jassy is telling Wall Street, suppliers and rivals one thing: the era of Amazon meekly overpaying for other people’s chips, networks and robots is over.

Behind the polite corporate phrasing sits a stark bet: $200 billion of capex in 2026 to build a vertically integrated AI and infrastructure empire — from custom silicon to satellites and warehouse robots. If he’s right, Amazon becomes the industrial backbone of the AI age. If he’s wrong, this is how an AI bubble turns into a very expensive hangover.

This piece unpacks what Jassy is really saying about Nvidia, Intel, Starlink and OpenAI — and what it means for the rest of us.

The news in brief

According to TechCrunch’s readout of Andy Jassy’s 2026 shareholder letter, Amazon is using its annual manifesto to quietly position itself against several of its biggest suppliers and partners.

On chips, Jassy argues that most AI workloads have so far run on Nvidia hardware, but claims a shift is underway toward Amazon’s own Trainium accelerators, hosted in AWS. He says demand for the latest Trainium generation is strong enough that current capacity is nearly sold out, and even the next generation — still roughly 18 months from availability — is largely pre-committed. He pegs Trainium’s annualized revenue at around $20 billion.

On CPUs, he highlights the success of Amazon’s Arm-based Graviton processors, stating that the vast majority of the top 1,000 EC2 customers already use them. Two large customers allegedly tried to reserve all Graviton capacity for 2026.

He also touts Amazon’s low-Earth-orbit satellite network, described as a Starlink rival called “Amazon Leo”, with early contracts from airlines, telecoms and even NASA. Jassy hints that Amazon may eventually sell robotics solutions based on data from its million-plus warehouse robots.

All of this underpins his justification for roughly $200 billion in capital spending this year, mostly for AWS data centers, partly backed by huge long-term cloud commitments from OpenAI and other customers.

Why this matters

Strip away the soft language and this letter is about power: who controls the bottlenecks of the AI economy.

For a decade, Nvidia and Intel have been the toll collectors on the road to the cloud. If you wanted serious compute, you ultimately paid them. Jassy is signaling that Amazon intends to move those toll booths in-house. Trainium and Graviton are not side projects; they are the mechanism to recapture margin, reduce dependency, and deepen customer lock-in.

If Trainium works as promised, AWS can offer cheaper, tightly optimized AI infrastructure while keeping more of the economics. That’s good for large cloud buyers and Amazon’s shareholders — and uncomfortable for Nvidia, which is increasingly surrounded by hyperscalers designing their own chips.

The same logic applies to Amazon Leo. SpaceX’s Starlink has been the default for low-latency connectivity in remote or conflict-prone regions. An Amazon-controlled satellite network gives AWS another way to bind customers into its stack — from cloud to edge to orbit.

The short-term winners are big AWS customers that can negotiate better price–performance on custom silicon and connectivity. The losers are traditional component suppliers whose biggest, richest customers are turning into direct competitors.

The real risk is execution. Custom chips, global satellite constellations and industrial robots are all capital-intensive, long-horizon bets. If AI demand normalizes or shifts faster than expected, Amazon could be left with overbuilt, underutilized infrastructure — and Wall Street’s patience is not infinite.

The bigger picture

Jassy’s letter fits into an unmistakable trend: the hyperscaler arms race to own every critical layer of the AI stack.

Google has spent years refining its Tensor Processing Units (TPUs). Microsoft has unveiled its own AI chips to reduce reliance on Nvidia. Meta is designing custom accelerators for its recommendation and AI workloads. Amazon was early with Graviton and is now doubling down with Trainium and dedicated networking.

This is structurally similar to what happened with smartphones a decade ago. Apple, Google and a few others gradually replaced off‑the‑shelf components with custom silicon, modems and operating systems. The result: better integration, higher margins, and brutal pressure on generic component vendors.

The scale is different this time. Public cloud and AI infrastructure capex is reaching levels reminiscent of the 1990s fiber build‑out and early-2000s data center spree. We know how those cycles ended: with spectacular write‑downs for some, and bargain infrastructure for the survivors.

At the same time, AI model companies like OpenAI, Anthropic and others are signing eye‑watering multi‑year commitments for compute capacity. Jassy leans on these deals to argue that Amazon’s capex is not speculative, but pre-sold. The uncomfortable reality: many of these startups are themselves unproven, heavily subsidized by investors, and operating in a policy environment that could change quickly.

Jassy brushes aside talk of an “AI bubble” for Amazon. History suggests otherwise: bubbles don’t look like bubbles from the inside. They look like “once-in-a-generation platform shifts” that “justify unprecedented investment”. He might still be right — but it’s precisely the kind of language we’ve heard before.

The European angle

From a European perspective, this letter should set off both excitement and alarm bells.

Excitement, because European enterprises, startups and public institutions continue to rely heavily on AWS. Cheaper, more efficient AI infrastructure based on Trainium and Graviton could lower barriers to entry for everything from biotech simulations to large-scale language models built in Europe.

Alarm, because this deepens Europe’s dependency on a handful of US hyperscalers that now control not just cloud software, but the physical chips and satellite links underneath it. While the EU debates digital sovereignty, the real leverage is shifting into layers where Europe has very limited presence.

The EU Chips Act, Gaia‑X, IRIS² (the planned EU satellite constellation) and the Digital Markets Act all aim, in different ways, to prevent a small number of foreign platforms from becoming unchallengeable. Yet Jassy’s vision is precisely that: AWS as a vertically integrated utility for compute, storage, networks, AI and even robotics.

Regulators in Brussels and national capitals will read this letter as a signal that Amazon is not just a marketplace or a cloud provider, but an infrastructure giant with growing systemic importance. Expect sharper questions around interoperability, data portability and fair access — especially if Trainium-only features or Leo-based connectivity become de facto standards for global AI workloads.

For European cloud challengers like OVHcloud, Deutsche Telekom, Orange or smaller national providers, Amazon’s move up the stack raises the bar yet again. Competing just on price or local compliance will not be enough if the hyperscalers can bundle custom chips, global satellites and robotized logistics into their offers.

Looking ahead

The next 24–36 months will test three of Jassy’s biggest claims.

First, that Trainium can win significant share of serious AI workloads. Watch what major SaaS vendors, game studios and financial institutions do, not what they say in joint press releases. If flagship AI products quietly migrate from Nvidia-heavy instances to Trainium-based ones, that’s real validation.

Second, that Amazon Leo can carve out space alongside Starlink. The early contracts Jassy mentions — airlines, telcos, space agencies — are important, but the harder part will be scaling coverage, reliability and regulatory approvals globally. For Europe in particular, how Leo coexists with IRIS² and national spectrum rules will be closely watched.

Third, that the AI demand curve really can justify $200 billion of capex in a single year. Investors should track not just AWS revenue growth, but utilization metrics, discount levels, and the health of big AI-native customers. If funding tightens or AI regulation bites harder than expected, some of those long-term commitments may be renegotiated.

On the robotics side, Amazon has the data exhaust of over a million robots in its warehouses — a treasure trove for building generalized industrial and maybe even consumer solutions. But entering the external robotics market would pit Amazon against highly specialized incumbents, in an environment with very different safety and liability expectations than its own facilities.

The long-term opportunity is clear: if Amazon’s vertically integrated stack becomes the default substrate for AI applications, its position could rival or surpass that of the old energy majors. The long-term risk is equally clear: overbuilding for a future that arrives more slowly, and under more regulation, than today’s hype suggests.

The bottom line

Jassy’s shareholder letter is less about comforting investors and more about warning suppliers and rivals: Amazon intends to own the critical infrastructure of the AI era, from chips to satellites to robots. The $200 billion capex plan might look reckless, but it’s strategically coherent — and brutally aggressive.

For users, this could mean cheaper, more powerful AI. For Europe, it deepens dependence on a US giant just as regulators push for sovereignty. The open question: will policymakers and competitors shape this future, or simply rent it from Seattle?