AI Watch: OpenAI-Anthropic War Escalates & Enterprise AI Data Risk

AI Watch | September 21, 2026

The AI Race Moves From Model Releases to Infrastructure, Enterprise Control and Economics

Market Signal

The AI Race Is No Longer About Which Model Is Smartest

Three developments this week made the shift unmistakable. The Financial Times reported that Big Tech has used residual-value guarantees to keep up to $300 billion in AI data-center and chip exposure off balance sheets—an enormous financing structure that raises questions about who ultimately carries the risk. Reuters reported that Anthropic is considering releasing another model ahead of a potential IPO, as competition with OpenAI's GPT-6 Astra intensifies and investors focus increasingly on enterprise spending and frontier-model economics. And Reuters further reported that Palantir, Nvidia and Booz Allen have curbed their own use of external AI models over concerns about proprietary data and model-provider data policies.

The common thread is not model capability. It is infrastructure, enterprise control, security and capital. The AI industry is maturing from a research race into an industrial competition—and the constraints that matter now are financial, operational and regulatory rather than purely technical.

The AI Infrastructure Bill Is Getting Bigger

AI infrastructure spending has reached a scale that no longer fits neatly inside corporate capital budgets. The FT's reporting on residual-value guarantees reveals the mechanism: Big Tech companies are guaranteeing the residual value of AI chips and data-center equipment to lenders, allowing financing partners to take on debt while the technology companies avoid recognizing the full liability on their balance sheets. The result is up to $300 billion in AI exposure that is economically real but financially obscured.

This structure matters because AI infrastructure has a shorter useful life than the financing assumes. GPUs depreciate quickly as new generations arrive. Data centers built for current-generation accelerators may need retrofitting within a few years. If residual values fall short of guarantees, the companies that provided them absorb the loss—and those losses are not currently visible in headline financials.

The capital requirement itself is staggering. Hyperscaler 2026 capital expenditure is projected to exceed $886.7 billion. TSMC has raised its own 2026 capital expenditure to $60–64 billion. Inference spending in AI-optimized infrastructure as a service is set to surpass training for the first time in 2026, reaching $23.3 billion compared to $19 billion.

Why this matters: The AI infrastructure bill is no longer just a technology question. It is a balance-sheet question, a credit-market question, and eventually a shareholder-return question. The companies that can fund infrastructure without obscuring risk will be better positioned when the cycle turns.

OpenAI vs. Anthropic Enters a New Phase

Reuters reported on September 19 that Anthropic is considering releasing another model ahead of a potential IPO. The timing is revealing. OpenAI's GPT-6 Astra has intensified competitive pressure, and investors are increasingly focused on enterprise AI spending, revenue growth and the economics of frontier-model development rather than benchmark leadership alone.

The enterprise market has become the primary battleground. Anthropic has leaned into enterprise deployments, regulated industries and sovereign AI arrangements—a strategy reinforced by the Cohere–Aleph Alpha merger announced September 16, which explicitly targeted regulated, deployable AI infrastructure running inside customer environments. OpenAI, meanwhile, has pursued consumer scale and a broad platform strategy spanning ChatGPT, enterprise agreements and custom silicon.

Open-weight models complicate the picture further. Meta's Llama family, Mistral's open models and DeepSeek's continued releases give enterprises an alternative to paying frontier-lab prices for every workload. For many enterprise tasks—document processing, classification, internal search—open-weight models running on customer infrastructure are sufficient, cheaper and more controllable.

Why this matters: The frontier-model competition is no longer decided by who has the smartest model. It is decided by enterprise adoption, developer usage, pricing power, revenue durability and IPO-readiness. The lab that wins the enterprise will not necessarily be the lab that tops the leaderboard.

Enterprise AI Has a Data-Control Problem

Reuters reported on September 14 that Palantir, Nvidia and Booz Allen have curbed their own use of external AI models over concerns about proprietary data and model-provider data policies. The Information first reported the internal pullback. These are among the most technically sophisticated organizations in the world—and they have concluded that the data-retention and training policies of major model providers do not meet their standards.

The issue is structural. Enterprise customers need zero-data-retention guarantees, cloud isolation, and clear contractual language about whether their inputs are used to train models. But model providers have incentives to use customer data for improvement—and the terms of service for many consumer and mid-tier enterprise tiers do not provide the isolation that regulated industries require.

The responses are taking several forms. Some enterprises are deploying open-weight models inside their own infrastructure. Others are negotiating dedicated instances with model providers. And a growing number are building internal AI platforms that abstract over multiple model providers—allowing them to route sensitive workloads to isolated environments while using frontier APIs for lower-risk tasks.

Why this matters: Data control is becoming a purchasing criterion, not a checkbox. Model providers that cannot guarantee isolation will lose regulated and security-conscious enterprise customers to open-weight alternatives and sovereign AI deployments. The Cohere–Aleph Alpha combination is a direct response to this demand.

Agentic AI Changes the Security Equation

As AI agents gain system access, tool use, and computer-use capabilities, the security surface expands dramatically. An agent that can read email, query databases, execute code, and interact with enterprise applications is effectively a non-human user with broad permissions. If compromised, it can cause damage at machine speed.

The enterprise response is still forming, but three patterns are emerging. First, identity and permissions: agents need scoped credentials aligned to the principle of least privilege, with the ability to revoke access instantly. Second, monitoring and auditability: every agent action must be logged, traceable and attributable to a specific agent and a specific authorizing human. Third, isolation and kill switches: agents that go off-script must be stoppable without disabling the entire platform.

The infrastructure vendors are racing to provide these controls. Airrived launched Agentic Observability for its Agentic OS, providing full traces from data ingestion to business outcome. Microsoft introduced Execution Container isolation in Agent 365 to block unsanctioned shadow agents. ServiceNow added a kill-switch capability in its AI Control Tower.

Why this matters: Agentic AI security is not a feature that can be added later. It is a prerequisite for enterprise deployment. The vendors that solve identity, permissions, auditability and isolation will win the enterprise agent market. Those that don't will be limited to low-risk experimentation.

What the AI Market Is Becoming

The developments this week connect across the full AI value chain:

Models → Compute → Data → Infrastructure → Enterprise Software → Security → Capital

Models: OpenAI, Anthropic, Google, Meta and open-weight challengers compete on capability, cost and enterprise fit. Model releases still matter, but they are no longer the primary source of competitive advantage.

Compute: GPUs, custom accelerators, advanced packaging and HBM determine how much AI compute is available and at what cost. The semiconductor supply chain is a binding constraint.

Data: Enterprise data control, zero-retention guarantees and proprietary corpora are becoming competitive moats. The companies that own the data can train, fine-tune and deploy models that others cannot.

Infrastructure: Data centers, networking, optics and power determine whether AI compute can scale. The $300 billion in residual-value guarantees shows how much capital is required—and how much is hidden.

Enterprise Software: Salesforce, ServiceNow, SAP, Oracle, Workday and Microsoft are building agent orchestration, governance and monetization layers. Gartner's $234 billion agentic arbitrage estimate frames the scale of the shift.

Security: Agent identity, permissions, observability, isolation and kill switches are becoming prerequisites for enterprise deployment. The security layer may determine which agent platforms win.

Capital: AI infrastructure financing, model-company IPOs, hyperscaler capital expenditure and strategic investments determine how fast the industry can grow—and who absorbs the risk if it slows.

What to Watch Next

Frontier-model launches: Anthropic's next model and OpenAI's GPT-6 Astra roadmap will shape enterprise purchasing decisions through year-end.
Enterprise AI spending: Watch for evidence of AI budget expansion versus consolidation—and for the first signs of ROI scrutiny affecting renewals.
AI infrastructure financing: The FT's $300 billion residual-value guarantee story will evolve as auditors, regulators and investors examine the off-balance-sheet structures.
Agent adoption: Whether enterprises move agents into production at scale—beyond pilots—will test the governance and security frameworks now being deployed.
AI security: Agent identity, permissions and observability vendors will compete to become the control plane for enterprise agent deployments.
Custom AI chips: Google TPU v7, Amazon Trainium 3, Microsoft Maia 200 and OpenAI's custom silicon program will test whether hyperscaler accelerators can scale beyond internal workloads.
Open-weight models: Llama, Mistral and DeepSeek releases will determine how much of the enterprise market remains addressable by frontier-lab APIs.
AI M&A and strategic investments: The Cohere–Aleph Alpha combination and the Salesforce–Fin acquisition signal continued consolidation across the AI stack.

The CODEW Analysis

The AI race is no longer about model releases. It is about infrastructure, enterprise control, security and capital—and the companies that master those layers will define the next phase of the industry.

This week's developments show how quickly the industry has matured. The $300 billion in residual-value guarantees reveals the scale of AI infrastructure financing—and the risk that comes with it. Anthropic's pre-IPO model considerations show that enterprise adoption and revenue durability now matter as much as benchmark performance. Palantir, Nvidia and Booz Allen's pullback from external models demonstrates that data control is becoming a purchasing criterion, not a checkbox. And the race to provide agent identity, permissions and observability shows that security is now a prerequisite for agent deployment.

The AI market is transitioning from a research competition into an industrial one. The winners will not be the companies with the smartest models. They will be the companies that control the infrastructure, the data, the governance layer and the capital required to scale. Model leadership is temporary. Infrastructure, enterprise trust and financial discipline are durable.

Sources

Financial Times — Big Tech uses guarantees to keep $300bn of AI exposure off balance sheets (Sept 2026)
Reuters — Anthropic considers releasing new AI model ahead of IPO, sources say (Sept 19, 2026)
Reuters — Palantir, Nvidia curb AI model use over data fears, The Information reports (Sept 14, 2026)
Reuters — Cohere, Aleph Alpha combine to target enterprise AI market (Sept 16, 2026)
MarketWatch — AI stocks are rebounding; analyst sees no spending slowdown
Gartner — $234 Billion in Enterprise Application Software Spend Is at Risk from Agentic AI
IDC — AI infrastructure spending projections
Airrived — Agentic Observability launch announcement
EyeOn.ai — Microsoft Agent 365 adds Execution Container isolation
ServiceNow — AI Control Tower governance documentation
Tom's Hardware — GPU shipment data
TSMC Q2 2026 earnings data

The CODEW Stat

Up to $300 billion in AI data-center and chip exposure is being kept off Big Tech balance sheets through residual-value guarantees. Hyperscaler 2026 capital expenditure exceeds $886.7 billion. Inference spending is set to surpass training for the first time, reaching $23.3 billion versus $19 billion. And Gartner estimates $234 billion in enterprise application spending is at risk from agentic AI through 2030. Stay updated, and thanks for visiting The CODEW. This is Erwin —see you in the next AI Watch stories!

The CODEW