Amazon
2026-06-30
AWS and Google Open Custom AI Chips for External Sales, ASIC Shipment Growth Surpasses GPU, TCO Inflection Point Reached
Summary
Key Takeaways
In Q2 2026, ASIC chips saw a flurry of strategic announcements. On June 18, AWS confirmed it is in talks to sell its custom Trainium chips to other companies' data centers, as confirmed by AWS AI business head Peter DeSantis. Anthropic has committed to deploying over 1 million Trainium chips, signing up to 5GW of chip capacity. Amazon CEO Andy Jassy hinted in his April shareholder letter that if the chip business were an independent entity, annual revenue would be about $50 billion; the internal chip division's annual revenue run rate has already exceeded $20 billion.
Google TPU announced a joint venture "TPU Cloud" with Blackstone in May, with Blackstone initially committing $5 billion (up to $25 billion with leverage). This marks the first large-scale commercial sale of TPU outside Google Cloud in its decade-long history. The project aims to bring online approximately 500MW of AI data center capacity by 2027.
TrendForce forecasts that custom AI chip shipments will grow 44.6% in 2026, while commercial GPU shipments will grow only 16.1%. This is the first time custom chip shipment growth has significantly surpassed general-purpose GPUs since the AI era began. Semianalysis and Bernstein estimate that ASIC TCO advantage over GPUs for large-scale inference ranges from 40% to 65%. Midjourney, after migrating to Google TPU, reduced its monthly compute cost from $2.1 million to $0.7 million.
Why It Matters
On the surface, AWS and Google opening custom chips meets demand for cost-effective AI compute, but the real intent is to encircle NVIDIA and weaken its GPU ecosystem dominance. By selling Trainium and TPU externally, they shift AI compute from standardized GPU commodities to vendor-locked custom platforms. Enterprises adopting Trainium or TPU get locked into AWS Neuron SDK or Google Cloud TPU software stacks, losing cross-platform portability.
The original text deliberately downplays ASIC's fatal lack of generality: Trainium and TPU are highly optimized for specific model architectures like Transformers but perform poorly on dynamic sparse models, multi-modal fusion, reinforcement learning workloads. Midjourney's TPU migration is a cherry-picked case—it perfectly fits TPU's matrix compute units. For AI labs frequently switching model architectures, ASIC's hardware inflexibility leads to massive hidden migration costs and performance penalties. Moreover, ASIC's longer iteration cycle (18-24 months) lags behind GPU's 12-month cadence, meaning customers may be locked into outdated hardware.
PRO Decision
【Vendors】 (Competitors: NVIDIA, AMD, Intel) should launch targeted attack marketing exposing ASIC's performance limitations on general AI workloads (multi-modal, reinforcement learning, dynamic graphs) and software ecosystem lock-in risks. NVIDIA should accelerate open-sourcing parts of CUDA (e.g., cuDNN, TensorRT) to lower migration costs, and introduce lower-cost inference GPUs (e.g., L40S series) to directly compete on TCO. AMD should emphasize ROCm's open-source nature and cross-platform compatibility as an anti-lock-in alternative.
【Enterprises】 CIOs and architects must perform zero-trust audits on ASIC external sales: demand benchmarks on non-optimized workloads (sparse models, multi-modal inference), estimate model migration costs (retraining, operator adaptation, performance tuning). Include software portability clauses in contracts to allow future multi-platform migration. Establish multi-vendor procurement strategies to avoid single-chip lock-in.
【Investors】 Reassess NVIDIA's moat depth. ASIC shipment growth surpassing GPU is a structural risk signal, but NVIDIA's CUDA ecosystem stickiness and generality advantage still support long-term value. Watch if NVIDIA counters via software open-sourcing or custom chip partnerships with cloud vendors. Also monitor AWS and Google's chip gross margins—external sales may depress cloud margins but expand TAM.
Get 3-5 key AI infrastructure signals weekly →
💬 Comments (0)