Models are not products. They are features. And features don’t command margins.
This is the hard truth that the April 28 OpenAI-AWS announcement exposes with surgical clarity. When Sam Altman sat down with AWS CEO Matt Garman to announce “Bedrock Managed Agents, powered by OpenAI,” he wasn’t making a distribution deal. He was signing the death certificate of the model-as-a-product business model.
Let me be more specific about timing, because the sequence matters. On April 27, Microsoft and OpenAI announced their amended partnership — ending Azure exclusivity, capping the revenue share, and scrapping the AGI clause that had hung over the relationship like a Damoclean sword. Twenty-four hours later, OpenAI’s models were available on a competing cloud platform. That is not a planned rollout. That is a pivot executed with the urgency of a company that realized its financial model had a clock on it.
The Stratechery interview dropped the same day, giving us direct quotes from both CEOs. Sam Altman described the Bedrock Managed Agents product as “a fundamentally new way for enterprises to deploy AI — not as an API call, but as an integrated part of their operational infrastructure.” Matt Garman, with the casual certainty of a man who has seen this movie before, said “This is how cloud platforms have always worked. The best technology wins when it’s accessible everywhere.”
Here’s what happened, stripped of the Silicon Valley spin: OpenAI’s crown jewel — the GPT model family that cost billions to train — is now a checkbox on an AWS console. It sits alongside Anthropic’s Claude, Meta’s Llama, and whatever open-weight model is trending this week. You pick your model like you pick your EC2 instance size. The UX is identical. The switching cost approaches zero.
That is not a victory lap. That is a surrender to commoditization.
The Infrastructure Embedding
Let’s be precise about what “Bedrock Managed Agents” actually is, because the press coverage is running the wrong headline. This is not “OpenAI models available on AWS” — that framing makes it sound like an API endpoint relocation. It is not.
Bedrock Managed Agents is an agentic orchestration layer that happens to use OpenAI models as its reasoning engine. I want to emphasize the architectural difference between this and what came before, because it is not subtle.
Previously, running OpenAI models on AWS required standing up an EC2 instance, installing the OpenAI SDK, managing your own API keys, building your own orchestration logic for multi-step tasks, handling authentication against your existing AWS resources, and implementing your own tool-calling infrastructure. It was possible — enterprises are creative — but it was bespoke. Every deployment was a custom integration project.
Bedrock Managed Agents eliminates all of that. The agent is a first-class AWS resource, managed through the same IAM policies, CloudFormation templates, and CloudWatch monitoring that govern every other AWS service. The OpenAI model becomes one configurable parameter in a managed resource definition. The enterprise does not need to know — or care — whether the reasoning engine is GPT-6, Claude Opus 5, or a fine-tuned Llama 4 variant. The agent just works. The value proposition is that your enterprise data already lives in AWS — S3 buckets, DynamoDB tables, Kinesis streams, Redshift warehouses. The managed agent connects to those data sources natively, executes multi-step workflows across your AWS environment, and uses an LLM to make decisions about what to do next.
The model — GPT, Claude, whatever — is the smallest part of this stack. It’s the CPU in a laptop. Necessary, but nobody buys a laptop for the CPU model number anymore.
This is the exact same playbook Amazon used to kill standalone database companies. Remember when Oracle was the most valuable part of the enterprise stack? Then AWS launched RDS, Aurora, DynamoDB — managed database services where the database engine became an invisible implementation detail. Oracle’s market share didn’t collapse overnight, but its strategic relevance did. The database became a feature of the cloud, not a product you bought separately.
The same dynamic is now happening to frontier LLMs. Bedrock Managed Agents is the Aurora moment for AI models.
Consider the technical architecture that makes this work. An agent running on Bedrock can spin up an AWS Lambda function to process a user’s request, query a vector database stored in Aurora PostgreSQL, fetch real-time data from Kinesis, call an external API via API Gateway, and write results back to S3 — all orchestrated through Step Functions. The LLM is invoked at decision points: “based on this user input and the context from these data sources, what should I do next?”
The model is a function call in a larger system. It is not the system. It is not the value. It is plumbing.
Consider a concrete example. A large retailer using Bedrock Managed Agents for inventory management might have an agent that:
- Receives a query: “Why is SKU-4482 out of stock at the Shanghai distribution center?”
- Queries DynamoDB for current inventory levels across all warehouses
- Fetches supply chain data from SAP (via an API Gateway integration)
- Checks shipping manifests stored in S3
- Runs a predictive model — a separate ML service, not the LLM — to estimate restock dates
- Generates a human-readable response using the LLM
The LLM is invoked at precisely two points: interpreting the initial natural language query and formatting the final response. Everything in between is standard AWS infrastructure. The enterprise is paying for the agent orchestration, not the intelligence. The model could be swapped without changing the workflow.
The $350 Billion Elephant
Now consider the valuation context that makes this move inevitable.
OpenAI just raised at a valuation that, by most estimates, exceeds $300 billion. Anthropic is at $350 billion. These are not technology company valuations — they are narrative valuations, and narratives have half-lives measured in funding rounds.
The infrastructure math is brutal. A single GPT-6 class training run consumes 10,000-15,000 H100-equivalent GPUs for 60-90 days. At prevailing rates, that’s $400-800 million per training run. And you need multiple runs because you’re iterating on architectures, safety fine-tuning, and domain-specific distillation simultaneously. OpenAI’s total compute burn across training, inference, and research is plausibly north of $5 billion annually and climbing.
The revenue side of the ledger doesn’t support this. OpenAI’s annualized run rate is estimated at $5-8 billion, heavily concentrated in enterprise API calls and ChatGPT subscriptions. Even at 80% gross margins (generous for inference-heavy workloads), the compute costs consume a staggering percentage of revenue — especially when you factor in that training costs are capitalized and amortized.
The fundamental equation is not solvable at the model layer. You cannot train frontier models at current scale and sell API access profitably. The unit economics are structurally negative at the frontier.
This is why OpenAI is doing two things simultaneously that look contradictory but are actually the same survival instinct:
Putting models on AWS Bedrock — surrendering model differentiation in exchange for distribution and, critically, access to AWS’s enterprise procurement cycles. Enterprises spend $100 billion+ annually on AWS. If OpenAI captures 1% of that as model-driven workload, that’s $1 billion in revenue that doesn’t require consumer acquisition costs.
Injecting ads into ChatGPT — monetizing the attention side of the business with contextual advertising. The Buchodi analysis published April 28 showed OpenAI’s ad platform serving Grubhub ads when users discussed Beijing travel, GetYourGuide for Great Wall tours, and Canva for productivity conversations. This is not a toy. The attribution infrastructure — four Fernet-encrypted tokens per ad, 30-day cookies, in-app webview tracking — is production-grade ad tech.
Two monetization strategies chasing the same gap: the gap between what frontier models cost and what enterprises will pay for raw intelligence.
The Agentic Capture
The truly strategic insight in the Bedrock Managed Agents announcement is not about models at all. It is about the agentic orchestration layer becoming the new permanent point of control.
Think about what happens when an enterprise deploys Bedrock Managed Agents throughout its organization. The agents have access to internal databases, document stores, communication channels, and operational APIs. They execute workflows that touch every part of the business — customer support triage, internal IT requests, compliance checks, data analysis, report generation.
Now, how hard is it to swap out the underlying model?
Answer: not very hard at all, because the model is just one component in a pipeline. AWS can swap GPT for Claude, Claude for Gemini, or any of them for a fine-tuned open-weight model with a configuration change. The switching cost that matters is in the agentic workflows, the data connections, the IAM permissions, the orchestration logic — all of which are deeply embedded in AWS.
This is cloud vendor lock-in 2.0. The first generation locked you into compute and storage. This generation locks you into agentic infrastructure that spans compute, data, and decision-making.
OpenAI is not giving up control by putting its models on AWS. It already lost control the moment it became clear that enterprises would never run their core business on a single model provider. What OpenAI is doing is choosing which layer of the stack to fight for. It has chosen the agent layer — through Bedrock Managed Agents — rather than the model layer.
Whether this choice was strategic or forced doesn’t matter. The outcome is the same: models are now a commodity input, and the value is in the orchestration.
The Anthropic parallel deserves examination here. Anthropic’s Claude is also available on AWS Bedrock, and has been for months. The difference is that Anthropic never had Azure exclusivity — they were always multi-cloud by design. Their $40 billion Google investment and $5 billion Amazon investment both came with cloud credit structures that ensured Anthropic would train on TPUs and infer on Trainium. Anthropic’s multi-cloud strategy was a hedge from day one. OpenAI’s is a correction.
This fundamental asymmetry matters because it means Anthropic has been architecting for multi-cloud compatibility since its inception. OpenAI has been optimizing for Azure’s single-stack efficiency. The engineering cost of retrofitting OpenAI’s stack for AWS — reworking training pipelines that assumed Azure-specific InfiniBand topologies, rebuilding inference infrastructure that expected ND-series GPU clusters — is non-trivial. These are not weekend projects. They are multi-quarter engineering efforts that will divert resources from model improvement to infrastructure compatibility.
The Parallel Histories
This pattern has played out before, twice, in ways that are instructive.
The Database Story (2005-2015): Oracle, IBM DB2, Microsoft SQL Server — enterprise databases were the most profitable software category in history. Gross margins above 80%, switching costs that measured in years, procurement cycles that lasted quarters. Then AWS launched RDS in 2009, Aurora in 2014, and DynamoDB in 2012. The database became a managed service. Oracle’s revenue still grew — the total addressable market expanded — but its margin structure collapsed. Database companies stopped being platform companies and became feature vendors.
The Smartphone Chip Story (2007-2017): When the iPhone launched, the ARM chip inside it was a marvel. By 2017, the A11 Bionic was doing machine learning inference on-device. But nobody bought an iPhone for the chip. The chip became an invisible enabler. Qualcomm, MediaTek, and Apple Silicon compete on specs that matter only to reviewers. The value is in the ecosystem — iOS, the App Store, iCloud.
AI models are following the same arc. They transition from “the product” to “the feature” to “the invisible infrastructure” in roughly three funding cycles. We are at the end of cycle two.
Strategic Implication
If models are becoming commodities — and the evidence strongly supports this — then the winners in AI will not be model companies. They will be:
The orchestration layer: Companies that build the agentic workflows, tool-calling infrastructure, and enterprise integrations that make models useful in context. AWS with Bedrock Managed Agents is the clearest play. Microsoft with Copilot. Google with Vertex AI Agent Builder.
The distribution layer: Companies that control the user interface where AI interactions happen. ChatGPT’s ad platform proves OpenAI sees this. But so do Google (Search, YouTube, Workspace), Microsoft (Office, Teams, Windows), and Meta (WhatsApp, Instagram, Facebook).
The hardware layer: Companies that manufacture the physical substrate — NVIDIA, Broadcom, TSMC. These are the only players with structural scarcity. You cannot software-around physics.
The model companies — OpenAI, Anthropic, even Google DeepMind to some extent — are caught in the middle. They produce the intelligence but cannot capture its full value because the intelligence is infinitely replicable at marginal cost once trained.
The only moat that matters in AI is physical: silicon, power, and data center real estate. Everything else is a lease that the cloud provider can revoke.
The Personal Verdict
I have been writing about the infrastructure debt trap for three weeks. The GPU debt treadmill. The CapEx avalanche. The sovereignty tax. Each article argued that the physical layer of AI — compute, power, silicon — would ultimately constrain and define the industry’s trajectory.
The OpenAI-AWS announcement confirms the opposite side of that thesis. If the physical layer constrains supply, the orchestration layer constrains value capture. OpenAI has made a rational decision: trade model exclusivity for access to enterprise infrastructure. But rational decisions can still be strategic defeats.
Here is the uncomfortable truth that nobody in the AI industry wants to say out loud: The most profitable part of AI might not be the intelligence at all. It might be the boring, unglamorous work of connecting that intelligence to existing enterprise systems — writing the API integrations, configuring the IAM roles, building the audit trails, managing the compliance certifications. The part that no VC demo day ever features.
AWS just announced that they own this layer. Microsoft already owns it through Azure and Office 365 Copilot. Google owns it through Workspace and Vertex AI. The model companies are renting space in someone else’s castle.
The margin structure tells the story with brutal clarity. AWS operates at roughly 30% operating margins on $100 billion+ in annual revenue. Microsoft’s commercial cloud runs at 45%+ margins. Google Cloud is approaching profitability after years of investment. These are infrastructure businesses with moats that deepen with every customer deployment — more data, more integrations, more locked-in workflows.
OpenAI’s gross margins on API inference are probably 50-70% after compute costs, if you exclude training amortization. Include training, and the number drops dramatically. Include the cost of frontier research — the endless pursuit of the next benchmark — and the unit economics turn negative. The model business is structurally less profitable than the infrastructure business, and the gap is widening as models grow larger and training costs accelerate.
The question is not whether OpenAI can survive this pivot. It almost certainly can — $300 billion valuations buy a lot of runway. The question is whether any company that defines itself primarily as a model provider can build a durable, high-margin business in a world where models are infrastructure.
The answer, based on everything we know about how infrastructure businesses work, is almost certainly no.
There is one scenario where this analysis is wrong. If OpenAI’s next model — GPT-6 or whatever comes after — is so decisively superior that enterprises are willing to pay a 5x premium for exclusive access to it, the model-as-a-product thesis gets a temporary reprieve. But the history of technology suggests this is unlikely. Intelligence appears to be following the same commoditization curve as compute, storage, and bandwidth. The gap between frontier models narrows with each generation. Open-weight models approach closed-model performance with a lag of months, not years.
OpenAI’s AWS pivot is not a mistake. It is an acknowledgment of physics — the physics of competitive markets, the physics of infrastructure economics, and the physics of silicon. Models are becoming a utility, and utilities do not command premium margins.
The age of model supremacy is over. The age of agentic infrastructure has begun. And the landlords — AWS, Azure, GCP — are already collecting rent.