Battleground 02 · October 2026 edition

Who Controls AI Compute?

Who supplies the compute AI runs on

Strategic Brief

September 11 – October 11, 2026 · 30-day retrospective · Evidence checked through October 11, 2026

The bigger trend

Compute advantage is shifting from contracted capacity to usable production.

Gigawatts under contract still dominate the headlines, but the question that matters to a buyer is how quickly capacity turns into working output and what each completed workload costs. This month the clearest signal came from a neocloud bundling the newest hardware, the developer tooling around it and new geographic supply into one offer.

Outside → In

What changed — Sense (verified)

CoreWeave, September 30 (announced at its Fully Connected conference). CoreWeave announced availability of NVIDIA Vera Rubin NVL72, with Cognition as the first customer running production workloads on it. Cognition's own engineers reported up to 4.8 times higher total token throughput on their SWE-2 inference workload versus a GB200 NVL72 baseline. The same day CoreWeave launched Forge, a development layer that combines Weights & Biases, OpenPipe post-training and marimo notebooks for training, inference, evaluation and agent development, available now in Free, Pro and Enterprise editions.

CoreWeave and AdaniConneX, October 7. CoreWeave plans 240 MW of capacity in Navi Mumbai across three 80 MW buildings. The first phase is expected in mid-2028. This is planned capacity, not installed capacity.

Context, not new this period: On September 17 CoreWeave disclosed about 4.2 GW of contracted power as of August 11, and said it had signed short-dated three-to-six-month contracts at about $40 million per MW, measured as annualized revenue divided by the power required.

What it means — Interpret (RAW PEC view)

The neoclouds are no longer competing on chips alone. By moving into the tooling where models and agents are built and evaluated, CoreWeave makes its capacity harder to leave: a team that trains, evaluates and observes agents inside Forge is a stickier customer than one renting GPUs. CoreWeave's short-dated contracts at higher prices are its own disclosed experience, not evidence of market-wide pricing power. These are one vendor's announcements; they are not evidence of market share.

What to consider — Position

When you compare compute suppliers, are you comparing contracted megawatts, capacity available on your timeline, or cost per completed workload? And how much of your development loop are you prepared to place with the same supplier?

Inside → Out — suggested enterprise actions

Suggested actions for readers, not a report of actions taken.

Who needs to align — Align

CFO and FP&A, infrastructure, engineering, sourcing and risk.

What the enterprise could execute — Execute

Model one production workload across two suppliers under identical assumptions for timeline, utilization and switching cost, including whether experiment tracking and evaluation tooling would sit with the compute provider.

How to measure and learn — Learn

Cost per completed workload, time from signature to usable capacity, utilization, and the effort required to move both the workload and its tooling.

Supporting competitive moves (this period)

CoreWeave · Build · Sep 30, 2026
Launched CoreWeave Forge, a development layer spanning training, inference, evaluation and agent development
Primary source · investors.coreweave.com
CoreWeave / Cognition · Partner · Sep 30, 2026
First customer running production workloads on NVIDIA Vera Rubin NVL72; Cognition reported up to 4.8x token throughput on its own inference workload versus GB200
Primary source · investors.coreweave.com
CoreWeave / AdaniConneX · Partner · Oct 7, 2026
Planned 240 MW in Navi Mumbai; first phase expected mid-2028
Primary source · investors.coreweave.com

RAW PEC Perspective

The compute contest is about who supplies the chips, software, networking and customer commitments that AI runs on.

The management question

Which strengths from the last inning will slow you down in the next?

Why now

Recorded summaries

Meta / CoreWeave: expanded $21 billion AI infrastructure agreement
Meta / Nebius: agreement worth up to $27 billion over five years
NVIDIA / Nebius: partnership to deploy more than 5 gigawatts of NVIDIA systems by end of 2030

RAW PEC Perspective

The stored moves are large capacity and investment agreements between model makers, silicon suppliers and AI clouds.

Attribution is as recorded and was not independently revalidated here.

Where this sits in the seven-layer stack

Contested layers: 2 Compute, cloud and security.

Adjacent layers: 1 Physical and energy foundations; 4 Models and intelligence.

Layer 7Business outcomes and value realization
Layer 6Workflow, agents and orchestration
Layer 5Domain platforms
Layer 4Models and intelligenceAdjacent
Layer 3Data, context and interoperability
Layer 2Compute, cloud and securityContested
Layer 1Physical and energy foundationsAdjacent

Contested control points

Layer 2, Compute, cloud and security: Which dependencies constrain continuity or replacement?

Contest dimensions

Chips
Software
Networking
Customer commitments

Decision lens

RAW PEC Perspective Compute control and economic stack, top to bottom

Customer commitmentsLong-term demand agreements that can support capacity investment
Neoclouds and hyperscaler cloudsWhere compute is delivered
Networking and softwareWhat links and programs the hardware
ChipsMerchant and custom silicon
Power and sitesThe physical base beneath compute

Text equivalent: An economic decision lens: customer commitments; clouds that deliver compute; networking and software; chips; power and sites. These are not substitute Canon layer names or numbers. The canonical seven-layer context is shown above.

A qualitative reading aid. It is not a ranking, readiness score, measurement or statement about any participant.

Participants

These are overlapping roles, not exclusive categories or rankings. Notes are as recorded, not independently revalidated.

Incumbents

NVIDIA (Incumbent): chips, CUDA software and networking; agreed to acquire Hugging Face
Intel (Incumbent): agreed to develop custom x86 CPUs for NVIDIA's platforms

Merchant and custom silicon

AMD (Challenger): merchant GPUs; 6-gigawatt deals with OpenAI and Meta
Broadcom (Challenger): custom accelerators; 10 gigawatts planned for OpenAI
Cerebras Systems (Challenger): wafer-scale chips; OpenAI agreement, then a 2026 IPO

Hyperscalers' own chips

Google (Challenger): seventh-generation TPUs offered to outside customers
AWS (Challenger): its own Trainium chips for its cloud

Neoclouds and AI factories

CoreWeave (Challenger): AI cloud; about 4.2 gigawatts of contracted power as of August 2026; nine of the ten leading model providers on its platform
Nebius (Challenger): AI cloud building gigawatt-scale AI factories; up to $27 billion, five-year agreement with Meta

Ecosystem moves

Each entry is shown as recorded, newest first. Type, date and status are kept separate. A reference appears only where one is recorded for that entry.

1 Build · 0 Buy · 9 Partner · 0 Compete

CoreWeave / AdaniConneX
planned 240 MW in Navi Mumbai; first phase expected mid-2028
Move type
Partner
Date
October 7, 2026
Date meaning
Date associated with the recorded entry (day precision); its event or publication basis is not specified here.
Status
Planned or exploring per the recorded summary; not a completed outcome.
Reference
Open entry-specific source
CoreWeave
launched CoreWeave Forge, a development layer spanning training, inference, evaluation and agent development
Move type
Build
Date
September 30, 2026
Date meaning
Date associated with the recorded entry (day precision); its event or publication basis is not specified here.
Status
The recorded summary describes an action or announcement; completion and operating outcomes are not determined by this status label.
Reference
Open entry-specific source
CoreWeave / Cognition
first customer running production workloads on NVIDIA Vera Rubin NVL72; Cognition reported up to 4.8x token throughput on its own inference workload versus GB200
Move type
Partner
Date
September 30, 2026
Date meaning
Date associated with the recorded entry (day precision); its event or publication basis is not specified here.
Status
The recorded summary describes an action or announcement; completion and operating outcomes are not determined by this status label.
Reference
Open entry-specific source
Meta / CoreWeave
expanded $21 billion AI infrastructure agreement
Move type
Partner
Date
April 2026
Date meaning
Date associated with the recorded entry (month precision); its event or publication basis is not specified here.
Status
The recorded summary describes an action or announcement; completion and operating outcomes are not determined by this status label.
Reference
No entry-specific reference is recorded.
Meta / Nebius
agreement worth up to $27 billion over five years
Move type
Partner
Date
March 2026
Date meaning
Date associated with the recorded entry (month precision); its event or publication basis is not specified here.
Status
The recorded summary describes an action or announcement; completion and operating outcomes are not determined by this status label.
Reference
No entry-specific reference is recorded.
NVIDIA / Nebius
partnership to deploy more than 5 gigawatts of NVIDIA systems by end of 2030
Move type
Partner
Date
March 2026
Date meaning
Recorded month associated with this entry; the deployment target is end of 2030.
Status
Planned deployment of more than 5 gigawatts by end of 2030, as described in the recorded summary; completion is not established by this entry.
Reference
No entry-specific reference is recorded.
OpenAI / Cerebras
750 megawatts of capacity, deployment planned
Move type
Partner
Date
January 2026
Date meaning
Date associated with the recorded entry (month precision); its event or publication basis is not specified here.
Status
Planned or exploring per the recorded summary; not a completed outcome.
Reference
No entry-specific reference is recorded.
OpenAI / AMD
6-gigawatt GPU agreement
Move type
Partner
Date
October 2025
Date meaning
Date associated with the recorded entry (month precision); its event or publication basis is not specified here.
Status
The recorded summary describes an action or announcement; completion and operating outcomes are not determined by this status label.
Reference
No entry-specific reference is recorded.
OpenAI / Broadcom
10 gigawatts of OpenAI-designed accelerators, deployment planned
Move type
Partner
Date
October 2025
Date meaning
Date associated with the recorded entry (month precision); its event or publication basis is not specified here.
Status
Planned or exploring per the recorded summary; not a completed outcome.
Reference
No entry-specific reference is recorded.
NVIDIA / Intel
$5 billion investment; Intel agreed to develop custom x86 CPUs for NVIDIA
Move type
Partner
Date
September 2025
Date meaning
Date associated with the recorded entry (month precision); its event or publication basis is not specified here.
Status
The recorded summary describes an action or announcement; completion and operating outcomes are not determined by this status label.
Reference
No entry-specific reference is recorded.

Evidence & RAW PEC Perspective

Recorded summaries and references

These summaries are not independently verified by this presentation. References are associated only with their mapped entries; a link does not establish support for every claim in a summary.

Meta / CoreWeave: expanded $21 billion AI infrastructure agreement