Google Cloud

Only hyperscaler with two competitive AI silicon lines — TPU (in-house, 7 generations) and NVIDIA — and the only one that hosts both Gemini (1P) and Anthropic Claude (large external customer) on the same infra.

1. Core Product / Service

Google Cloud Platform (GCP) is Alphabet's cloud business; the AI-relevant slice is Cloud TPU + GPU VMs (A3/A4) + Vertex AI managed-model layer + Gemini API.

Compute SKUs that matter for AI:

• Cloud TPU v6e (Trillium) — sixth-gen TPU, 4.7× peak compute / chip vs v5e, 2× HBM, deployed in 256-chip slices and pods scaling to ~9,000+ chips [2][3].
• Cloud TPU v7 (Ironwood) — announced Apr 2026; 5x compute of Trillium, optimized for inference at scale [7].
• A3 (H100) / A3 Mega / A3 Ultra — 8× H100 with 3.2 Tbps Jupiter-network fabric.
• A4 (B200) — 8× B200 Blackwell.
• A4X (GB200 NVL72) — rack-scale GB200, GA late 2025.
• Vertex AI Model Garden — managed catalog (Gemini 1P, Claude, Llama, Mistral, partner models).
• Gemini API — direct 1P API, runs on TPU primarily.

Capacity strategy: TPU is the cost/perf weapon (Google designs/owns the silicon), NVIDIA covers customers with CUDA-locked workloads. Both TPU and NVIDIA scaled massively for Gemini training and the Anthropic relationship.

2. Target Users & Pain Points

• Anthropic — uses Google Cloud TPU as a primary training and inference platform; Oct 2025 expansion announced up to ~1M TPU chips and >1 GW of capacity [5][6].
• Frontier labs beyond Anthropic — DeepMind (captive), various external research labs use TPU for cost.
• Enterprise — Snap, Spotify, PayPal, Wayfair, Wendy's standardize on Vertex / Gemini for managed AI.
• Search-adjacent customers wanting low-latency Gemini integration (Gemini 2.5 Pro / Gemini 3 Pro Preview).

Pain points addressed: TPU price/perf below NVIDIA on supported workloads (JAX/PyTorch-XLA), tight Workspace + Search integration, only cloud where Claude is first-class alongside the 1P model.

3. Competitive Landscape

GCP's edge: TPU silicon stack is 4 generations ahead of any other hyperscaler's custom ASIC; only cloud running both 1P frontier model (Gemini) and external lab (Anthropic) at scale. Disadvantage: smaller installed base, less enterprise sales muscle than AWS/Azure, perceived "developer-only" reputation.

4. Unique Observations

• A3 H100 pricing: a3-highgpu-8g (8× H100 80GB) on-demand list $88.49/instance·hour ($11.06/H100·hour); 1-year CUD $55.74/instance·hour ($6.97/H100·hour); 3-year CUD $39.82/instance·hour ($4.98/H100·hour) [1]. A3 Mega / A3 Ultra (H200) list higher; A4 (B200) introductory list $110/instance·hour ($13.75/B200·hour) for short reservations, deeper discounts under Spot/CUD.
• TPU pricing (Trillium v6e): ~$2.70/chip·hour list on-demand, ~$1.89/chip·hour 1-year CUD, ~$1.22/chip·hour 3-year CUD [2]. TPU v5p (training-tier) at ~$4.20/chip·hour on-demand. TPU per-FLOP is 2-3× cheaper than equivalent H100 spend for workloads that can use it — the central reason Anthropic moved a large share of training onto TPU.
• TPU + NVIDIA dual capacity: Google does not disclose split, but Anthropic deal alone implies massive TPU ramp — ~1M chips of new capacity dedicated [5][6]. NVIDIA fleet on Google Cloud also large but secondary in strategic narrative.
• Anthropic partnership is the most significant external-lab cloud relationship for Google. Anthropic uses both AWS Trainium (primary, via the Amazon $8B investment) and Google TPU/NVIDIA (large training + inference partner). Google has invested ~$3B+ in Anthropic cumulatively (2023 reports of ~$2B + 2024 reports of ~$1B+ extension; full numbers undisclosed). The 2025 expansion to up to 1M TPU chips makes Anthropic the second-largest single TPU consumer after Google's own DeepMind/Gemini training [5][6].
• AI revenue share: Google Cloud Q1 2026 revenue $13.42B (+30% YoY) [4]; operating income $2.83B. Sundar Pichai consistently calls AI "the dominant driver" of Cloud growth in earnings; the precise AI share is not disclosed, but analyst estimates put AI-attributable Cloud revenue at $4–$6B annualized run-rate. Cloud is still ~13% of Alphabet total revenue ($90B Q1 2026).
• Customer concentration: not publicly disclosed; Anthropic is widely understood to be the largest external AI compute customer; internal captive load (Search AI Overviews, Gemini, YouTube ML, DeepMind) likely exceeds external AI revenue. Vertex's other anchor accounts (Snap, PayPal, Wendy's) are publicly named but at smaller spend.

5. Financials / Funding

• Parent: Alphabet (NASDAQ: GOOGL); Google Cloud is reportable segment.
• Q1 2026 Google Cloud revenue: $13.42B (+30% YoY); operating income $2.83B (~21% margin) [4].
• FY2025 Google Cloud revenue: ~$48B (+30% YoY) [4].
• AI-attributable revenue: not disclosed; Pichai cites AI as "primary growth driver" — analyst estimates $5B+ annualized.
• Capex: Alphabet FY2026 capex guidance ~$95B, "vast majority" AI infrastructure (Pichai/Ruth Porat commentary).
• Anthropic investment: ~$3B+ cumulative (combined Google + Google Ventures rounds 2023–2024).
• TPU program: 7 generations shipped (v1 → Ironwood v7); production volumes not disclosed but Anthropic's 1M-chip commitment alone is record-setting [5][6][7].

6. People & Relationships

• Alphabet CEO: Sundar Pichai.
• Google Cloud CEO: Thomas Kurian.
• DeepMind: Demis Hassabis (Gemini training captive customer of GCP TPU).
• Anchor external AI partner: Anthropic (~1M TPU expansion 2025; ~$3B+ Google equity).
• NVIDIA: critical supplier for A3/A4/A4X.
• Other AI partners on Vertex: Mistral, Meta (Llama), Cohere, Stability.
• Competitors: aws, microsoft-azure, oracle-cloud, coreweave.

Sources

[1] https://cloud.google.com/compute/gpus-pricing (2026-05-10) [2] https://cloud.google.com/tpu/docs/v6e (2026-05-10) [3] https://cloud.google.com/blog/products/ai-machine-learning/introducing-trillium-6th-generation-tpu (2026-05-10) [4] https://abc.xyz/assets/investor/2026Q1_alphabet_earnings_release.pdf (2026-05-10) [5] https://www.anthropic.com/news/expanding-our-use-of-google-cloud (2026-05-10) [6] https://www.reuters.com/technology/artificial-intelligence/anthropic-google-tpu-expansion-2025-10-23/ (2026-05-10) [7] https://blog.google/products/google-cloud/ironwood-tpu-age-of-inference/ (2026-05-10)