Power, inference efficiency, and AI control planes reshape cloud competition

By DripPublished

The gist

This week, cloud competition shifted from raw model access to control of scarce infrastructure, inference economics, agent governance, and trusted execution.

This week’s developments

AI Capacity Competition Shifts to Power Control

Google’s next-gen TPU rollout, Deutsche Telekom’s nearly €250 million Munich AI factory expansion to 20 MW, and KT Cloud’s 50-MW AI-dedicated site in South Korea show the AI race moving from chip procurement to control of power, cooling, land, and accelerator access. That matters because scarce advantage is now being locked into bundled infrastructure, not just compute orders: frontier compute agreements are roughly 46% custom silicon, 39% Nvidia, and 15% AMD, while Cerebras has already secured about 700 MW through end-2027. For operators, speed to power is becoming the product; for vendors and investors, value is concentrating in proprietary silicon ecosystems and power-secured AI infrastructure.

How do we secure power and infrastructure before competitors do?

If you operate in this industry

  • AI capacity is now won by power, land, and cooling—not just chips.
  • Secure power-backed sites and accelerator supply early, or lose AI growth to better-capitalized rivals with faster time-to-power.

Sources

If you sell into this industry

  • Demand is shifting to bundled AI infrastructure, not standalone compute.
  • Align roadmap and GTM to power-secured, silicon-tied deals; budget is moving to vendors that can deliver full-stack capacity.

Sources

  • AI’s Physical Bottlenecks: GPU Markets and the Chip Investment Thesis — Cognitive Revolution "How AI Changes Everything", October 22, 2026

    Explores turnkey GPU cluster management, partnerships with power and data-center owners, and forward markets for lending accelerator capacity.

  • A Tale of Two Computes — Token Dispatch, October 11, 2026

    Explains how concentrated supply, fixed-price contracts, and strategic financing shape compute demand and market positioning.

  • Barely Possible - 2026-11-10 — Barely Possible, November 9, 2026

    Anthropic’s turbine purchase reveals how power scarcity is driving buyers toward self-supplied energy—and shifting infrastructure and operating-cost risks.

If you invest in this industry

  • Value is concentrating in power-secured AI infrastructure and silicon ecosystems.
  • Favor firms with locked-in power, land, and custom silicon access; pure compute and point plays face margin and access pressure.

Sources

Nvidia Turns Inference Efficiency Into the Next AI Cost War

Nvidia’s March 2025 Dynamo release pushed the optimization fight into inference execution: it splits LLM prefill and decode across GPUs, reuses KV caches, and spills older cache blocks to CPU memory, local storage, or object storage so tokens are not recomputed. Paired with TensorRT-LLM techniques such as kernel fusion, quantization, chunked prefill, multiblock attention, and speculative decoding, Nvidia says H200 throughput can rise by up to 3.6x. The hardware roadmap is being framed the same way: Blackwell GB300 is positioned at up to 35% lower token cost and 50x higher throughput per megawatt than Hopper for DeepSeek-R1, while Rubin, announced January 5, 2026, targets further KV-cache sharing and up to 10x lower inference cost per token than Blackwell.

That shifts the market from into measurable execution efficiency. Enterprises are now operationalizing token-level attribution, GPU utilization tracking, commitment planning, and pre-call policy enforcement through Finout, IBM Cloudability, CloudZero, Revenium, FinOpsly, Cast AI, and others. RouteLLM has shown up to 85% lower MT-Bench cost at roughly 95% of GPT-4 quality, while prompt caching has cut API costs by 41% to 80% across 500-plus agent sessions. Cost-per-task is becoming an architecture decision, and vendors that can prove savings in production will pressure clouds and GPU stacks to compete on efficiency outcomes, not capacity alone.

Who captures margin as inference efficiency becomes the new battleground?

If you operate in this industry

  • Inference cost is now a competitive lever, not just a cloud bill.
  • Track token-level unit economics and GPU utilization; redesign workloads and commitments around cost-per-task, or rivals will undercut you on AI economics.

Sources

If you sell into this industry

  • Buyers want provable inference savings, not generic AI performance.
  • Shift roadmap and GTM toward measurable token-cost reduction, cache reuse, and utilization gains; efficiency proof now closes deals.

Sources

If you invest in this industry

  • AI value is shifting to efficiency winners, not just capacity sellers.
  • Favor infra and software that can prove lower inference cost per token; point tools without hard savings evidence face margin and multiple pressure.

Sources

AI Control Planes and Converged Data Stacks Become the New Cloud Battleground

AWS, Microsoft, and Google have each moved agent infrastructure up the stack: Bedrock AgentCore, Foundry Agent Service with Entra Agent ID, and Google’s Gemini Enterprise Agent Platform with Agent Runtime all center on managed execution, identity, and tool access for agents. Microsoft also expanded Agent 365 into a cross-platform control plane for governance, visibility, resilience, and cost optimization, while AvePoint’s AgentPulse adds ownership mapping, attestations, and remediation workflows. The message is clear: cloud vendors are no longer selling models alone; they are competing to own the operating layer where agents are governed and monetized.

LangChain’s reported $300 million Series C at a $3 billion valuation on Nov. 13 reinforces that the orchestration layer between models, data, and applications is becoming a strategic control point. In parallel, EnterpriseDB is pushing a converged PostgreSQL + ClickHouse stack, using its Postgres Analytics Accelerator to publish live data to Apache Iceberg so transactional and analytical engines can query a shared catalog with less ETL and duplication. With enterprise support, security hardening, Kubernetes operations, and deployment across on-prem, private/public cloud, hybrid, and air-gapped environments, the competitive prize is clear: vendors that control agent governance and open data movement will capture more of the cloud value chain.

Where will control-plane value accrue, and how should we position?

If you operate in this industry

  • Cloud platforms are racing to own your agent control layer.
  • Treat agent governance, identity, and tool access as strategic infrastructure; avoid lock-in by keeping orchestration and data portability open.

Sources

  • D2DO315: Running Agents Securely at Scale — The Everything Feed - All Packet Pushers Pods, October 7, 2026

    Explore agent identity standards, Kubernetes scaling, token costs, and tradeoffs between local models and cloud providers.

  • D2DO315: Running Agents Securely at Scale — The Fat Pipe - All Packet Pushers Pods, October 7, 2026

    Explores agent identity standards, Kubernetes operations, hosting trade-offs, and cost considerations for deploying agents at scale.

  • D2DO315: Running Agents Securely at Scale — The Everything Feed - All Packet Pushers Pods, October 7, 2026

    Learn about agent identity provisioning, Kubernetes operations, token costs, and trade-offs between cloud-hosted and local model deployment.

If you sell into this industry

  • Governance and orchestration are now the buying center, not models alone.
  • Shift roadmap and GTM toward managed execution, auditability, and cross-cloud control; point features without control-plane depth will get squeezed.

Sources

If you invest in this industry

  • Value is moving to control planes and open data stacks, not raw models.
  • Favor platform owners and infra enablers; underwrite point tools cautiously as bundling and shared-data architectures compress standalone pricing.

Sources

AMD and Google Cloud Put Confidential AI Inside the TPU Path

On November 9, AMD and Google Cloud pushed the story further: Google Cloud’s new TPU 8tx runs on 5th-generation AMD EPYC processors with AMD SEV/Infinity Guard, and Google says it is the first cloud TPU deployment with hardware-enforced confidential computing. The key shift is that sensitive prompts, model weights, and data are protected across the EPYC host and TPU path for inference, fine-tuning, and reinforcement learning, extending the trusted execution boundary from CPU-only systems into accelerator-based AI.

That announcement lands alongside a broader platform race over default AI isolation. AWS’s Nitro Sandboxes for Amazon EKS gives each Kubernetes pod a lightweight VM boundary through Kubernetes RuntimeClass, while Google’s Agent Compute Engine runs each agent in a hardware-isolated microVM with its own kernel, ephemeral identity, controlled networking, and hardware-rooted attestation and credential handling. Rafay, Fortanix, and KubeVirt are operationalizing the same model for Kubernetes and mixed VM-container governance. For operators, this is the next step beyond confidential containers: less friction for sensitive AI on shared infrastructure, and more pressure to treat integrated attestation, key release, and accelerator-aware trust orchestration as baseline cloud capabilities.

Where does confidential AI create the next moat and margin?

If you operate in this industry

  • Confidential AI is moving into the accelerator path, not just the CPU.
  • Treat attestation, key release, and accelerator-aware trust as core infra; shared AI capacity now needs hardware-backed isolation to stay competitive.

Sources

If you sell into this industry

  • Hardware-enforced AI isolation is becoming a cloud buying criterion.
  • Shift roadmap and GTM toward attestation, microVMs, and accelerator trust orchestration; buyers will favor native cloud-aligned controls over add-ons.

Sources

If you invest in this industry

  • Confidential computing is expanding from niche to AI infrastructure default.
  • This validates spend in platform security and trusted AI infra; watch for winners that own the control plane, not just point tools around it.

Sources

Stay ahead in Cloud Computing

Get the weekly Cloud Computing brief in your inbox — the developments, what they mean by vantage, and what to do next.