Your AI Stack Is Becoming a Supply Chain

For most of the generative-AI era, teams treated the model as the dependency. Pick OpenAI, Anthropic, Google, or an open model. Connect an API. Build the product.

This week made that architecture look increasingly incomplete. OpenAI said it intends to stop supplying its models to Cursor following Cursor’s acquisition by SpaceX. Anthropic introduced a standard designed to let agents operate physical hardware from robotic arms to microscopes. NVIDIA began shipping a CPU built specifically around the orchestration work agents perform between model calls. ChatGPT added the ability to pull multiple Google identities into one conversation. And Google DeepMind pushed its latest video model toward production workflows with separate economics for drafting and final output.

Individually, these are different announcements. The pattern underneath them is more important: an AI product is no longer one application sitting on top of one model. It is becoming a supply chain of models, contracts, compute, permissions, standards, tools, and data. PMs and founders need to start managing it that way.

1. OpenAI Just Showed That Model Access Can Become a Business Risk

On August 28, OpenAI announced that it intends to wind down the agreement under which it supplies models to Cursor, proposing a cutoff date of November 12, 2026. The decision follows Cursor’s acquisition by SpaceX earlier this month. OpenAI says its agreement contains a change-of-control provision and that it is exercising that option because it is not confident SpaceX would use OpenAI technology within its contractual terms. That is OpenAI’s stated rationale; Cursor itself is not accused in the announcement of violating the agreement.

Ignore the corporate drama for a moment. The product lesson is bigger. A model dependency can disappear because of something that has nothing to do with API reliability, model quality, engineering, price, or latency. A change in ownership can do it. So can a change in contract terms, regulation, or safety policy. Vendor availability itself is now part of your product-risk model.

For every critical AI workflow, define: Primary model → fallback model → acceptable degradation → migration time → contractual exposure.

This does not mean every startup needs five models running in production. It means the abstraction between your product logic and your model vendor needs to be intentional. The PM question is no longer only, ‘Which model performs best?’ It is also, ‘How expensive would it be to replace this model next Friday?’ Model portability is becoming business continuity.

2. Anthropic Is Standardizing the Jump From Software to Physical Machines

On August 27, Anthropic opened a research preview of the Model Hardware Standard, or MHS. MHS is a shared specification intended to let AI agents operate programmable physical equipment including microscopes, liquid handlers, robotic arms, and scientific instruments. Anthropic says integrating equipment in laboratories or manufacturing environments can currently require weeks or months of bespoke integration work. MHS attempts to replace much of that work with standardized drivers and common commands such as read and write.

The system is explicitly model-agnostic and can expose devices through standard protocols including MCP. We spent the last year standardizing how an agent accesses software tools. Now the same abstraction is beginning to reach hardware. Anthropic describes early work with organizations including Genentech, where Claude coordinated a liquid handler, robotic arm, and microplate reader during a protein-assay proof of concept.

Standards compress integration cost. And when integration cost collapses, entirely new product categories become economically viable. The interesting startup opportunity may not be building the smartest robot. It may be building the workflow, permissions, orchestration, monitoring, or vertical application layer above standardized machine interfaces. PMs should watch standards carefully. Standards often look boring right before they become distribution.

3. NVIDIA Built a CPU Around the Work Agents Do Between Model Calls

AI infrastructure discussion usually starts and ends with GPUs. Agentic systems are making that framing obsolete. On August 27, NVIDIA said its Vera CPU is now beginning to ship at scale, with AWS receiving its first Vera CPU server and Vera Rubin GPU. NVIDIA says Vera has 88 custom Olympus cores, 1.2 TB/s of memory bandwidth, and can deliver up to 1.8× faster per-core performance on certain agentic workloads compared with the reference systems it uses for comparison.

Why build a CPU specifically around agents? Because a real agent does much more than inference. Between model calls, the system may execute Python, search databases, retrieve context, run a sandbox, call APIs, process files, coordinate multiple tools, maintain state, and run simulations. NVIDIA explicitly positions Vera around this orchestration layer.

A better production equation is: model inference + CPU orchestration + retrieval + tool calls + sandbox execution + storage + networking + retries + monitoring + human review = cost per completed outcome.

Stop calculating AI cost as tokens × model price. That metric becomes increasingly misleading for agents. Your infrastructure decisions can eventually affect your unit economics as much as your model choice. For high-volume AI products, architecture is product strategy.

4. ChatGPT Can Now Combine Multiple Google Identities in One Conversation

On August 28, OpenAI added the ability to connect multiple Google accounts for Gmail, Google Calendar, and Google Contacts to ChatGPT. That means a user can, for example, connect personal and work calendars and ask ChatGPT to reason across them in the same conversation. The capability is rolling out globally across supported Plus, Pro, Business, and Enterprise plans.

At first glance, this looks like a convenience feature. It points to a much larger product problem. The more useful AI assistants become, the more identities they touch. A person does not have one context: they have personal email, work email, multiple calendars, different Slack workspaces, GitHub accounts, CRM access, client environments, banking identities, and health accounts. The same user may have different permissions inside every one.

Identity scope should become visible in the UX of agentic products: Account → workspace → permission → action.

As AI products connect to more systems, hidden context becomes dangerous context. Permissions are not backend plumbing anymore. Permissions are product design.

5. Google Is Separating AI Draft Economics From Production Economics

On August 27, Google DeepMind released Gemini Omni 1.1 Flash for developers with new controls aimed at production video workflows. The model can extend scenes, interpolate between specified first and last frames, reference video inputs, and upscale final generations to 1080p or 4K. The feature I find most interesting is much less flashy: Google now offers 360p preview generation designed for rapid iteration. The company says these drafts can be generated up to 60% faster and at roughly one-third the cost of standard 720p generation.

That is product thinking, not just model thinking. A professional creative workflow does not need every experiment rendered at maximum quality. It needs cheap draft → evaluate → refine → expensive final render. Build quality tiers into expensive AI workflows: Draft mode for cheap and fast iteration, Review mode for better quality at moderate cost, Production mode for maximum quality, and Escalation mode for a premium model or human review. Do not spend frontier-model economics on every intermediate step.

The Thread Connecting All Five

OpenAI and Cursor show that model access is contractual. Anthropic’s MHS shows that interfaces are becoming standardized. NVIDIA Vera shows that agent infrastructure extends beyond GPUs. ChatGPT’s multi-account support shows that identity boundaries are becoming a core AI product surface. Gemini Omni shows that production AI needs staged economics, not one quality level for everything.

These do not look like the same story. They are. The AI stack is becoming a supply chain. And supply chains have failure modes. A vendor disappears. A contract changes. A connector loses permission. A hardware dependency becomes expensive. An identity boundary is crossed incorrectly. A new standard wins. A model becomes uneconomical. A provider changes its policy.

When the industry was mostly building chatbots, many of these risks were tolerable. If the chatbot failed, the user regenerated the answer. Agents are different. An agent might update Salesforce, deploy code, move money, schedule a patient, operate laboratory equipment, run an ad campaign, or send an email under your name. Now every dependency underneath the agent becomes part of the product.

The AI Dependency Map

For every critical workflow, document the dependencies that determine whether the product can continue operating safely and economically.

• Model — What intelligence does it depend on?
• Fallback — What replaces it?
• Data — Which systems supply context?
• Identity — Under whose permissions does it operate?
• Tools — Which external systems can it modify?
• Infrastructure — What compute and runtime does it require?
• Contracts — What happens if a vendor relationship changes?
• Cost ceiling — At what price does the workflow stop making sense?
• Human escalation — Who owns the failure?
• Portability — How long would migration take?

That may sound like infrastructure work. It is increasingly product work. Because the product the customer experiences is not your prompt. It is the reliability of the entire chain behind it.

Build for capability. Design for substitution. Operate for failure.

Stack & Ship is a daily newsletter covering AI, product, startups, and operations — written for builders and operators who want the practical side, not the hype. Published by Manvendra Kumar — Product Manager and builder behind Mopshy AI, CareBow, OnliGrow, and ProductJarvis.com.