AWS History and Trainium's AI Future; OpenAI's Microsoft Deal

Summary

• AWS has reaccelerated enough to reopen, but not settle, Amazon’s AI-infrastructure thesis. AWS revenue grew 28%, its fastest pace since 2022, while Amazon revenue rose 17% to $181.5 billion and net profit jumped 77% to $30.3 billion, partly reflecting pretax income from Amazon’s investment in Anthropic. Both figures beat analyst estimates. Jassy attributed part of the AWS growth to AI-agent builders wanting their agents stored where their existing cloud services and data already reside.
• Amazon’s defining advantage is a structurally lower cost to serve commodity demand. Thompson’s simplified example has interchangeable providers spending $10, $9, $8 and $7 per widget: if the market-clearing price is $10, the lowest-cost producer makes $3 and retains room to cut prices while weaker competitors exit. Amazon is “the anti-Apple,” investing for cost superiority rather than premium differentiation.
• AWS combines custom silicon with service abstraction so customers need not see where Amazon saves money. Early Graviton processors “sucked,” but Amazon could probably use them behind managed services such as Redshift; Nitro separately offloaded networking and hypervisor work, letting AWS fit “like, 20% more virtual machines” onto expensive Intel capacity.
• Feature breadth turns AWS’s cost advantage into lock-in and pricing power. Customers begin with portable compute and storage, then adopt one convenient proprietary API after another until “you fast-forward and you’re totally locked in.” Margins once expected near 1–2% were roughly 17–19% when Thompson’s “AWS IPO” exposed them, and he recalled that they are now in “the 30s—or maybe even higher.”
• The bearish case remains credible for large-scale AI training because NVIDIA’s architecture increasingly treats the entire data center as one GPU. That demands tightly linked chips, racks and NVIDIA networking, clashing with AWS’s proprietary-networking playbook; Thompson said Amazon has never been a big training player and named Microsoft, Oracle, Elon Musk and xAI as players building dedicated data centers for it.
• Inference is the potential reversal because efficiency, utilization and cost matter more than giant horizontally connected clusters. Distillation can keep a model within one chip, while CPUs orchestrate jobs and batch sizes keep GPUs full—“that’s where you’re actually selling.” The preview ends before Thompson answers whether the AWS AI story looks better or whether Trainium has risen from the dead.

Deep dive

1. AWS reaccelerated, but the market signal was noisy

• Sharp’s Wall Street Journal excerpt put AWS growth at 28%, its fastest since 2022. Amazon revenue rose 17% to $181.5 billion, while net profit climbed 77% to $30.3 billion, partly reflecting pretax income from Amazon’s investment in Anthropic. Both figures beat analyst estimates.
• Jassy tied the surge to AWS’s cloud edge and aggressive data-center investment, while explaining that AI-agent builders often want their agents in the same cloud as their existing services and data.
• The excerpt had shares up more than 4% after hours, but Thompson thought Amazon had since “ended up down so far today.”

2. Amazon wins commodities by making the same thing cheaper

• Thompson’s framework, under simplified assumptions about substitutable supply and scalable demand: with interchangeable widgets costing four providers $10, $9, $8 and $7, the market-clearing price can reflect the highest-cost marginal supplier. The $7 producer then earns a sustainable $3.
• If prices fall, the highest-cost supplier exits, supply contracts and pricing can recover; the lowest-cost operator therefore has both resilience and competitive power. “The lower your cost structure in the industry, the better you are.”
• His memorable comparison was Amazon as “the anti-Apple,” meant positively. Apple sustains differentiation through hardware, exclusive software, developer ecosystems, network effects and brand; Amazon spends for years building cost advantages in commodity markets, repeating the playbook across retail and cloud.

3. AWS turns custom silicon and feature breadth into lock-in

• Early Graviton processors “sucked,” Thompson said, but managed services hid the underlying hardware. Customers buying Redshift received a database service, while Amazon could probably power it with cheaper Graviton capacity and improve the chips over time.
• Nitro handled networking, system management and hypervisor work beside the main processor—the “janitorial aspects of the server.” Thompson estimated AWS could fit “like, 20% more virtual machines” onto one Intel chip, structurally lowering cost versus Microsoft.
• AWS also won by arriving first and relentlessly adding features. Its 80/20 problem is that every customer wants Amazon to remove everyone else’s complexity while retaining “this one thing that I need.”
• Customers promise to remain portable, then use one convenient AWS API after another: “You fast-forward and you’re totally locked in.” That supports pricing power, while the cost base still enables startup credits, multi-year commitments and discounts Thompson said could reach 80%.
• The economics surprised investors: Thompson recalled AWS margins of roughly 17–19% when his “AWS IPO” exposed them, against expectations of 1–2%; he said they are now in “the 30s—or maybe even higher.”

4. Large-scale training exposes Amazon’s networking disadvantage

• The SemiAnalysis critique Thompson revisited was not wrong: Amazon optimized around proprietary networking, while leading AI systems expanded from individual GPUs to racks and then linked data centers. Jensen Huang’s framing—“the entire data center as a GPU”—requires NVIDIA networking and a full-system approach that undercuts AWS’s traditional strategy.
• Thompson called that concern “all true in terms of training,” which requires horizontal scaling and extremely low latency between chips and systems. Training consumed, by his rough recollection, roughly 60% of global chips for a long period, including several years after ChatGPT; he named Microsoft, Oracle, Elon Musk and xAI among the players building dedicated data centers, not Amazon.

5. Inference could return AI to Amazon’s home field

• Thompson’s conditional turn was “when and if inference came along.” Inference generally tries to keep work within one chip, including through model distillation, rather than coordinating enormous horizontal clusters; CPU orchestration, batch size and keeping GPUs occupied become more important.
• Sharp’s inference—endorsed by Thompson—was that this market should be more commoditized, with efficiency mattering more in inference than it does in training, where performance is not the only factor.
• The economic endpoint matters: in theory, if all the training is worthwhile, its share of compute should shrink because training produces a model whose value is realized through inference. That could favor AWS’s cost playbook, but Sharp’s repeated question—“Has Trainium risen from the dead?”—remains unanswered when the free preview cuts off.