Broadcom announced something today called Tanzu Platform Agent Foundations. The name reads like an apology, three nouns stacked to soften the ambition, but the bet underneath it is direct. AI agents, the company is saying, will not live in the cloud as floating intelligence. They will run on something, and that something will look a lot like the enterprise plumbing already paid for. The announcement promises to bring PaaS simplicity to enterprise agent deployment on VMware Cloud Foundation, with the audit trails and provisioning conventions that operations teams already understand.
That is the whole argument of the week, and it is worth stating before the evidence rather than after it. The value in this cycle will not collect in the model. It will collect in the substrate: the place where the work runs, where the tooling lives, and where developer hours accumulate until leaving becomes unthinkable. Every significant announcement this week is a different party placing a bet on that same square, and most of them are not describing it in those terms because describing it plainly would make the stakes obvious.
Broadcom, Snowflake, and the Race to Be the Floor
Snowflake, in the same news cycle, announced an expanded AI partnership, and the stock soared in extended trading on the news. Different vendor, same instinct. Snowflake wants agents to look like queries running against a warehouse, which is to say a thing the buyer already trusts, already provisions, and already pays for. Not a new procurement category with a new budget fight attached, but a new line item on an old invoice. That is a much easier sale than intelligence, and it is a much stickier one, because the invoice is the relationship.
You can read those two announcements as a coincidence of scheduling. They are better read as the same announcement filed twice. Both companies are arguing that the agent is a workload, not a product, and a workload has to be scheduled, isolated, observed, and billed by somebody. Whoever does that job is the one with the durable position, and neither company needs to build a competitive model to hold it.
Four Business Models, One Question
A piece in TechTimes this week catalogued four business models emerging for AI agents: open-source infrastructure, token distribution, SaaS, and acquisition. The article frames this as a strategic split, a matter of picking your approach and committing. It is better understood as a single question asked four different ways, and the question is whose floor agents stand on. Open-source infrastructure says the floor is a public commons. Token distribution says the floor is a chain. SaaS says the floor is a vendor’s runtime. Acquisition says the floor is whoever already owns the room. Each path is the same wager about where the value will eventually collect, placed by parties with different assets to defend.
The model itself is not where it collects, and we have watched that movie already. The state-of-the-art frontier model in March is a price war in November. The agent framework everyone raves about gets rewritten over a long weekend by a teenager in Lisbon. The protocol that promises to be the standard is forked twice before the documentation is finished. What stays is the place where all of it runs, and the reason it stays is boring: migration costs are real, tooling is sticky, and an organization will tolerate a mediocre floor far longer than it will tolerate rebuilding one.
This is also why a small room matters more than its attendance suggests. Ben Hylak noted this week that OpenAI is throwing an autoresearch hackathon on Saturday with Raindrop and Modal. A few hundred builders, and most of the projects forgotten by July. But the patterns invented in rooms like that, the conventions, the file layouts, and the way agents are expected to talk to data, end up baked into infrastructure that millions of developers touch two years later. The hackathon is not where products get built. It is where the grain of the wood gets decided, and the grain is what everything downstream has to cut along.
Why Ethereum Keeps Winning Onchain Finance
In crypto the same dynamic is playing out, just further along the curve. Grayscale’s Zach Pandl pointed this week at Ethereum’s continued dominance in onchain finance metrics: value locked, stablecoin issuance, and settlement volume. Set that against the TechTimes framing and the friction is instructive. If the four business models really were four strategies, the technically superior chain would take share from the incumbent on the merits. There are technically superior chains. There are cheaper chains. There are faster chains. Ethereum keeps winning the share of finance that lives onchain anyway, for exactly the same reason VMware will keep capturing enterprise AI workloads despite having no special claim to being the best place to run an agent.
The mechanism is not mysterious and it is not sentimental. The substrate is where the tooling lives. The tooling is where the developer hours go. The developer hours are where the moat compounds, one integration at a time, until the cost of moving exceeds the benefit of the better option. That is not lock-in in the contractual sense. Nobody is forbidden from leaving. It is something more durable than a contract, because it does not require enforcement.
Structure Decides What You Build
Donella H. Meadows wrote in Thinking in Systems that “structure determines what behaviors are latent in the system.” She was writing about ecosystems and feedback loops, not software. Read against the current tooling moment it lands like a diagnosis.
Consider an enterprise running Erlang, a language engineered for fault-tolerant, concurrent workloads, that is now weighing a migration to Java. Not because Java is architecturally superior for what the team is building. Not because the engineers prefer it. Because the AI coding assistants work better with Java, since the training corpus had more Java in it. The structure of the toolchain has made certain behaviors cheap and others expensive, and the organization is quietly bending to fit that structure without ever making a deliberate decision to do so. There was no meeting where anyone chose worse concurrency semantics. There was only a long series of small moments in which one path had help and the other did not.
That is precisely what Meadows is pointing at. The behaviors available to you, how you build and what you build and which tradeoffs surface naturally and which never get considered at all, are not determined by your intentions or your roadmap. They are determined by the structure you are operating inside. The substrate is the structure, and structure is already making choices on your behalf, upstream of any meeting you will ever have. This is the dynamic Broadcom is betting on, the dynamic Snowflake is betting on, and the dynamic Ethereum has been quietly winning for years. Whoever owns the floor where the work happens shapes what work gets done, not by locking anyone in, but by making certain moves natural and others unthinkable.
The Wrong Question, Asked Loudly
The popular question right now is which AI agent will be most powerful: which framework, which model, which orchestration layer. It is the wrong question, and it is the kind of question that gets asked at high volume while the real game is played underneath the table. The right question, the one Broadcom and Snowflake and Ethereum and the hackathon organizers are all answering in their own dialects, is simpler and quieter. Whose ground will all of this rest on?
Models will swap. Frameworks will churn. Agents will be rewritten every six weeks for the next two years, and nobody will remember the names of three-quarters of them. The ground does not move. The ground compounds. Watch the floors, not the dancers.
Leave a Reply