A WW3.PRESS SPECIAL PROJECT01 — INTELLIGENCE & POWER

An interactive essay

The war
beneath
the models.

Different ambitions.
One connected machine.

Read the argument
Conceptual glass-and-porcelain island connecting power generation, data-centre towers, chip fabrication and communications.
Concept illustration
Select a layer. Follow the connections.

OpenAI · Anthropic · Google · DeepSeek · Qwen

A connected system, not a league table.
THE PREMISE

Call it an arms race. Call it an industrial revolution.
The important question is who controls what everyone else needs.

01 / The contenders

Rivals. Suppliers.
Sometimes both.

The names on the chat window are only the visible layer. Choose a lab to see one part of the network.

GPT / ChatGPT

OpenAI

Infrastructure partnerships

Stargate's original announcement names SoftBank, OpenAI, Oracle and MGX as initial equity funders. [1]

A model developer's reach also depends on financing, cloud capacity and how people access its tools.

An announced buildout is not a completed buildout.

One documented connection
OpenAI
Stargate: announced partners
Oracle · SoftBank · MGX
Claude

Anthropic

Competing and cooperating

Anthropic's April 2026 statement identifies Amazon as its primary cloud and training partner while expanding its Google/Broadcom relationship. [2]

Supplier diversity can change a lab's dependence on any one provider. It also makes a simple two-sided rivalry misleading.

The new TPU agreement anticipates capacity beginning in 2027.

One documented connection
Anthropic
Stated compute relationships
AWS · Google · NVIDIA
Gemini / Google Cloud

Google

Models and the machinery

Google develops Gemini and offers its own Tensor Processing Units through Google Cloud. [3][4][2]

A company can compete in the model layer while supplying the infrastructure another lab uses.

This connection does not make the companies one team.

One documented connection
Google Cloud
TPU supplier relationship
Anthropic
DeepSeek model family

DeepSeek

A different distribution route

The R1 repository publishes reasoning-model weights and distinguishes the terms applying to its distilled variants. [5]

Releasing weights lets other people build around a model. That distributes adoption differently from relying entirely on one hosted interface.

Open weights do not automatically mean full training transparency.

One documented connection
DeepSeek-R1
Published model weights
Developers · deployers
Alibaba Cloud

Qwen

An ecosystem, not a country

Qwen3's repository documents public weights and multiple ways to run and deploy that model family. [6]

Qwen and DeepSeek are distinct organizations. Treating all Chinese models as a single actor obscures meaningful differences.

Check each model's terms; one release is not a universal licence.

One documented connection
Qwen3
Release and deployment ecosystem
Developers · deployers

A selected set of organizations, not a complete industry map. Connections describe the cited records; they are not live capacity measurements.

Analysis · 21 September 2026 · WW3.press

The AI race is a contest
for the infrastructure
of power.

The struggle over advanced AI reaches beyond a contest between chatbots. It is a contest over the machinery, access and institutions through which intelligence becomes power.

“World War III” is the frame of this essay, not a claim that a world war between technology companies has begun. Commercial rivalry, espionage and armed conflict are different things.

01. A refinery for intelligence

The object above is imaginary. The dependency it represents is not. A model reaches a person through a chain of computing systems, infrastructure and organizational decisions. Calling the result “the cloud” makes the physical part easy to forget.

In its April 2025 report, the IEA projected that global data-centre electricity use would more than double to around 945 TWh in 2030. That is a forecast for the whole data-centre sector, with AI an important growth driver—not a measurement of AI's current demand. The report also identifies grid-connection delays as a constraint. [7]

The refinery is a useful visual analogy because it asks us to look at inputs, conversion and distribution. It has a limit: intelligence is not a barrel of oil. Models can be copied, improved and deployed in very different ways. The illustration is a map of questions, not an engineering model.

02. The rivals share pipes

The easiest story puts each company behind its own flag. The evidence is more interesting. Anthropic says Claude uses AWS Trainium, Google TPUs and NVIDIA GPUs. In April 2026 it announced additional Google/Broadcom capacity expected to start coming online in 2027. [2] Google develops Gemini as well as its own accelerators. [3][4]

OpenAI's original Stargate announcement, in January 2025, likewise described a network of funders and technology partners. Its proposed $500 billion investment over four years was a stated intention—not a receipt for completed infrastructure. [1]

That is the tension at the heart of this project. Competition does not eliminate dependence. It can intensify it. A lab can seek independence from one supplier while tying itself more closely to another. Our interpretation is that control of the connections may matter as much as a temporary lead in model performance.

03. Beyond two teams

“Chinese models” is a category, not a single competitor. DeepSeek and Alibaba's Qwen have different organizations and model families behind them. The DeepSeek-R1 and Qwen3 repositories document examples of publicly released weights and ways for other people to build on those releases. [5][6]

Those examples complicate a closed contest between a few hosted products. A model can spread through developers, local deployments and downstream services. Open weights do not guarantee that training data, every training step or every later release is equally open. Each release needs its own examination.

This is also why the five names in our explorer are a starting point. They are not a claim that the rest of the industry, academic research or the people using these systems have ceased to matter.

04. Where conflict becomes real

The metaphor becomes serious when AI systems are used in security operations. In November 2025, Anthropic reported an espionage campaign misusing Claude Code and assessed the operator as Chinese state-sponsored. That attribution and the account of the operation come from Anthropic; this essay has not independently verified its incident evidence. The company also acknowledged model errors and described defensive uses of the same capabilities. [8]

That report is not evidence that Chinese model developers directed the operation, or that AI labs are attacking their commercial rivals. Connecting those dots without evidence would turn an explainer into a conspiracy story.

Cyber operations, proxy conflicts and conventional warfare are potential contexts in which AI may matter. They do not belong in one undifferentiated bucket. A commercial partnership, a government decision and a military act each require their own sources, actors and standards of proof.

05. A finish line worth questioning

“The race to AGI” sounds as if somebody installed a finish line. The research picture is less tidy. The Levels of AGI framework separates performance and generality, and considers autonomy and deployment risk as additional dimensions. It is a proposal for making comparisons clearer, not a universal referee. [9]

Our position is that a useful public discussion should ask what a system can do, under which conditions, at what cost, with which permissions and with what degree of reliability. A dramatic label cannot answer those questions.

There is a bigger struggle here than a product launch. It concerns whose tools become indispensable, who can deny access, and who gets to define an acceptable risk. Calling that a world war is provocative. Making its dependencies visible is the useful part.

06. What to watch

Watch for operating capacity rather than announced capacity. Watch the terms under which people can use and modify a model. Watch the difference between a capability demonstration and reliable deployment. And when a security claim appears, ask who observed it, what evidence is public and what remains uncertain.

The answer is not a countdown to an inevitable catastrophe. It is a habit of following the connections—and checking whether the story survives contact with its sources.

Independent analysis.
Evidence you can follow.

03 / The evidence

Open the sources.
Test the argument.

Reviewed 21 September 2026. Company statements are attributed. Projections and editorial interpretations are labeled.

1OpenAIThe Stargate announcementCompany announcement

Establishes the proposed infrastructure venture and named partners. An investment intention is not money already spent or capacity delivered.

21 January 2025Read original source
2AnthropicGoogle and Broadcom compute agreementCompany announcement

Documents Anthropic's stated hardware mix and a capacity agreement expected from 2027. Future capacity remains a company expectation.

6 April 2026Read original source
3Google DeepMindGemini model familyProduct documentation

Identifies Google's model family. This page is not independent evidence of comparative superiority.

Reviewed 21 September 2026Read original source
4Google CloudIntroduction to Cloud TPUTechnical documentation

Explains Google's machine-learning accelerators. Hardware documentation does not establish a lab's available capacity.

Reviewed 21 September 2026Read original source
5DeepSeekDeepSeek-R1 repositoryModel release

An example of released reasoning-model weights and documented terms, including distinctions for distilled models. Not a current-model ranking.

2025 release · reviewed 21 September 2026Read original source
6Qwen / Alibaba CloudQwen3 repositoryModel release

Documents publicly released weights and deployment options for this family. Do not generalize its terms to every Qwen release.

2025 release · reviewed 21 September 2026Read original source
7International Energy AgencyEnergy and AIResearch / projection

The 945 TWh figure is a 2030 projection for all data centres, not today's consumption or AI-only demand.

10 April 2025Read original source
8AnthropicReported AI-assisted espionage campaignVendor incident report

Describes Anthropic's investigation and attribution assessment. This essay has not independently verified the underlying incident.

13 November 2025 · corrected 14 NovemberRead original source
9Morris et al. / ICMLLevels of AGIResearch paper

A proposed framework separating performance, generality and autonomy; not a certification that a particular system has reached AGI.

2024 paper · revised 24 September 2025Read original source

A project from WW3.press

Understand the world
behind the interface.

Explore the reporting. Follow a source. Bring your own questions.