Amazon and Google didn’t just poke the Nvidia bear this week—they walked right into its cave with new silicon and big customers in tow. So if you’re eyeing the dip, do you reach for NVDA…or AMZN?
At AWS re:Invent (Dec. 2), Amazon launched Trainium3, a new in-house AI accelerator. Amazon says Trn3-powered UltraServers pack 144 chips, deliver ~4.4× the compute of its prior generation, and use ~4× less energy—claims paired with customer case studies citing up to 50% lower training/inference costs.
Google, meanwhile, keeps scaling its TPU line. Its v5p hardware—now generally available—clusters into massive pods (docs list up to 8,960 chips), and Google is courting external buyers. Meta is reportedly in talks to rent TPUs in 2026 and buy them starting 2027—an explicit challenge to Nvidia’s dominance.
And the supply chain is morphing. Marvell agreed to buy Celestial AI to bring photonic interconnects into its portfolio—an arms race ingredient for faster, cooler data centers.
This is about price, power, and availability. AWS pitches Trainium3 as a lower-cost, lower-energy alternative at scale—useful if GPUs are pricey or back-ordered. Even more intriguing: Amazon says its next chip will incorporate Nvidia’s NVLink Fusion fabric, suggesting future systems may blend best-of-both rather than force a binary choice.
Google’s push is similar: make TPUs a credible, cheaper path for large training jobs—and sell the capacity, not just use it internally. If Meta signs on, that’s real volume leaving the GPU queue and validating TPUs for third-party workloads.
None of this means Nvidia is retreating. Far from it. The company just posted $57B in quarterly revenue with $51B from data center, and still commands an estimated 80–90% share in training silicon. Shares did fall ~10–12% in November as investors weighed rising competition—but the platform (CUDA, libraries, ecosystem) remains the moat to beat.
The NVDA case: You’re betting that Nvidia’s full-stack advantage (chips + networking + software) keeps hyperscalers loyal even as they diversify. Watch: Blackwell ramp, networking attach rates, and whether large customers slow GPU orders as alternatives mature. If data-center growth and margins stay fat, the November wobble looks like digestion, not derailment.
The AMZN case: You’re betting that AWS converts Trainium3 from a press release into utilization—Bedrock and Trn3 UltraServers filled with paying workloads—and that price/performance plus power savings attract customers who don’t need CUDA. Key tells: third-party wins beyond early partners, sustained “up to 50%” cost claims in real deployments, and any sign that Trn3 eases GPU scarcity for AWS clients. Also note: Amazon’s stock is only ~6% up YTD—lagging several mega-caps—so the bar may be lower if execution improves.
The Google swing factor: If a Meta-TPU deal closes, chip spend that once auto-flowed to GPUs could diversify. Pair that with v5p pod scale and improving pricing, and Google becomes not just a software rival but a hardware supplier to peers.
A realist’s take: This isn’t a winner-take-all market, more like “winner-takes-most, others take plenty.” Hyperscalers will mix and match for cost, power, and model portability. For investors, the scoreboard over the next 3–4 quarters is simple:
• Nvidia: data-center revenue/margins and software stickiness.
• Amazon: Trainium3 adoption and customer savings that show up in AWS growth.
• Google: external TPU deals that convert headlines into backlog.
Bottom line: competition is finally real at the high end of AI compute. Whether you prefer NVDA’s entrenched platform or AMZN’s cost-driven challenger story depends on what you think matters more in 2026—ecosystem or economics.
Disclaimer: Investing carries risk. This is not financial advice. The above content should not be regarded as an offer, recommendation, or solicitation on acquiring or disposing of any financial products, any associated discussions, comments, or posts by author or other users should not be considered as such either. It is solely for general information purpose only, which does not consider your own investment objectives, financial situations or needs. TTM assumes no responsibility or warranty for the accuracy and completeness of the information, investors should do their own research and may seek professional advice before investing.
