Six AI cloud categories
A new practical guide breaks the 2026 AI cloud landscape into six distinct categories — GPU/accelerator clouds, managed ML platforms, serverless AI endpoints, agent orchestration, hybrid/edge AI clouds, and open-source orchestration frameworks — forcing platform teams to pick trade-offs across performance, cost, and governance argued. That taxonomy makes benchmarking and platform selection a core engineering competency, not just a procurement decision.
Janakiram MSV published the taxonomy piece on The New Stack on Mar 15, 2026, mapping vendor trade-offs and providing a comparison table and evaluation framework. thenewstack.io Deloitte’s TMT Predictions estimated that inference will account for roughly two‑thirds of AI compute by 2026, and projected the market for inference‑optimized chips to exceed US$50 billion while data‑center AI chips top US$200 billion. deloitte.com NVIDIA’s rack‑scale GB300 NVL72 configuration bundles 72 Blackwell Ultra GPUs and 36 Grace CPUs per rack and, in MLPerf Inference v5.1 submissions, the GB300 NVL72 delivered about a 45% DeepSeek‑R1 throughput gain versus GB200 systems. nvidia.com OpenAI’s compute footprint has become multi‑vendor: OpenAI and AWS announced a multi‑year infrastructure partnership reported at roughly US$38 billion in announced commitments, while CoreWeave expanded its OpenAI agreement by up to US$6.5 billion, bringing their cumulative deal value to about US$22.4 billion. openai.com Neocloud players are scaling fast: Lambda raised a US$480 million Series D to expand GPU capacity and achieve NVIDIA Exemplar Cloud status, and Crusoe recently reached a reported US$10 billion valuation while launching modular “Spark” factories and edge zones to accelerate regional AI capacity. pitchbook.com MLCommons’ MLPerf Inference v5.1 added the DeepSeek‑R1 reasoning benchmark and published results from 27 submitters, underscoring why The New Stack’s comparison checklist advocates workload‑specific benchmarking and SLA‑matched metrics for procurement and platform engineering decisions. mlcommons.org
