AI Startups Are Betting On What The Big Platforms Cannot Credibly Sell
Forbes
Güney Yıldız Oct 21, 2026, 06:40am EDT
Summary
OpenAI, Google, and Workday are pitching enterprise buyers to control AI agents, while leaving liability with humans in the loop. This leaves independent AI startups questioning how to compete against platforms with near-limitless compute. Industry leaders at Amsterdam's HumanX conference suggest startups survive by holding positions platforms cannot occupy, such as specialized data access, portable cross-model memory, multi-cloud neutrality, and domain-specific workflows. Meanwhile, venture-backed AI companies have actively acquired other startups.
As OpenAI, Google and Workday bid to run corporate AI agents, founders in Amsterdam point to four positions the giants struggle to occupy: portable memory, cross-cloud neutrality, domain workflow QA and decision-first data.
In the space of two weeks, OpenAI, Google and Workday each made a pitch to corporate buyers bidding to become the control point where company AI agents are chosen, governed and paid for. None of the three is eager to answer who pays when an agent errs, a question where vendors still park the liability with a human in the loop. Yet their expansion leaves every independent AI startup facing an equally blunt question: what do you sell that the platform cannot absorb on its next roadmap?
At HumanX Amsterdam last month, I put a version of that question to founders and executives across the floor. If the frontier labs have near-limitless compute and modular intelligence, why sell tokens cheaply instead of taking over every software category themselves?
My own suspicion going in was that the honest answer leaves startups with very little room. Once the model becomes the most replaceable part of an enterprise system, the giants can bundle the surrounding software and put the remaining vendors on a marketplace shelf. The founders I pressed in Amsterdam disagreed. Their answers point to a sharper divide: the difference between selling a feature, which a platform will copy or buy, and holding a position a platform cannot credibly occupy.
Why Would A Frontier Lab Bother?
Stefan Weitz, HumanX’s co-founder and chief executive, hears the question every week. “I’m also an investor in venture funds, so I get pitched 30 companies a week,” he said at the conference’s media briefing. He recalled his years at Microsoft: “Would you get run over by Microsoft? Yeah, probably.” Yet large companies leave whole markets alone because the addressable market is too small, the workflow is too specialized, or they lack the sales force to deploy it. “I actually think we’ll see more startups,” he said. What separates them “would be not necessarily the model itself but the data that they have access to and can use.”
Kostas Chalkias, co-founder and chief cryptographer at Mysten Labs and a former lead cryptographer at Meta, pointed to internal bureaucracy and distribution. “Every decision for a new product might take you one year to go through all of the, you know, hierarchy of people fighting funding,” he said. “How do you win against Uber if you don’t have distribution?” Moving into vertical industries is “a distraction for them if they do it,” Chalkias argued. Instead, he expects the labs to buy rather than build: “They are going to acquire probably some companies that do that instead of them doing that.”
The deal numbers back his reading. Venture-backed AI companies bought 195 other AI startups through September 29, more than in all of 2025, according to Crunchbase. OpenAI led with 10 acquisitions this year, while Anthropic made five. Even the startups on the platforms’ shelves are consolidating: legal AI firm Harvey has made four acquisitions this year and customer-service specialist Sierra has made three. This month alone, Salesforce agreed to acquire AI analytics startup Loops, confidential-computing startup Opaque announced its team was joining Anthropic in a license-and-hire deal, and Decagon, valued at $4.5 billion in a March tender, was reported to have hired a bank to field takeover interest, though ServiceNow denied involvement.
Buying is still a way of eating. For a founder, an acquisition is an exit. For a chief information officer deciding whether a supplier will exist in three years, it is a counterparty risk to price.
Who Owns The Company’s Memory When The Model Changes?
If a company can swap one model for another in an afternoon, the real lock-in moves to everything its agents have learned about the business. Chalkias noted three reasons companies switch models: a better model arrives, a cheaper one does the job, or a regulator or employer restricts which systems can be used.
He pointed to this summer’s interruption, between June 12 and July 1, when access to Anthropic’s Fable and Mythos models was suspended. “There was a point in time that we were stopped from accessing Fable and Mythos. And imagine at that time, what do I have to do?” Chalkias said. In most enterprise setups, switching models means leaving accumulated context behind. “You’re not able to go and get your data and move it to another LLM.”
Mysten’s proposed answer is Walrus Memory, built on the Walrus data platform and anchored to Sui, which lets a user grant different AI agents a delegate key to the same encrypted memory so switching models does not mean starting again. “Because you control the keys, you encrypt the memory based on your keys that you own... Nobody can stop you,” Chalkias said.
Mysten says more than 5,000 agents across 4,800 owners have been onboarded since launching in the second quarter, recording more than 209,000 stored memories and 4.5 million successful retrievals by October 20. Among owners whose first write occurred on or before September 19, the company says 72.9% retrieved at least one memory 30 or more days later. Those are company figures inferred from encrypted read-write patterns rather than an independent audit of paying enterprise deployments, and an export key alone cannot guarantee that a new model will interpret old context the same way. Still, the structural point holds: no single model maker has a commercial reason to make walking away with your company’s memory effortless.
Three Things A Platform Struggles To Sell Itself
Beyond portable memory, the founders in Amsterdam staked out three operational positions where a giant platform’s own business model gets in its way.
Neutrality across rival clouds. Roeland Pater of Dutch governance startup Deeploy described its software as “a layer that sits under or over all systems that are there within your organization,” adding, “we are agnostic that way.” Deeploy can also be deployed on a customer’s own infrastructure “for sovereignty.” Most large enterprises run across more than one cloud and model provider. A bank using both Microsoft and Google is unlikely to let one vendor’s gateway audit the other’s agents, and the company selling the agents makes an awkward auditor of its own work. Mysten has leaned into the same gap, announcing a collaboration with Google Cloud on October 6 to build Verifiable Agent Arbiter, an evidence layer that decouples audit logs from any single provider. Deeploy similarly leaves policy judgment with the buyer. “We don’t, let’s say, censor any output,” Pater said. “We basically give companies the technology to monitor that output themselves.” That outside check matters more as the real security risk shifts from what agents know to what they can do.
Domain depth and workflow QA. May Habib, CEO of Writer, argued that horizontal platforms underestimate how brittle enterprise workflows are. “The harness decides how the model works,” she said. “And if you don’t have a harness that is really domain specific, you are going to reteach AI constantly.” Her sharper critique was about connectors. Many vendors that advertise enterprise integrations, she said, have plugged into a third-party connector that works only some of the time, leaving agents “failing silently” when an authentication token expires across thousands of seats. “The value is in the workflow and the QA of the workflow,” Habib said, pointing to Writer’s Enterprise Brain, which encodes shared team memory and compliance rules directly into regulated workflows. Google’s own Gemini Enterprise launch concedes part of that ground: its legal workflow version remains in preview and relies on startups Harvey and Onit as partners.
Starting from the decision, not the data lake. Bianca Anghelina, co-founder and CEO of Aily Labs, has watched companies spend years trying to clean their data before letting AI touch a decision. “You will never end this project,” she said. Because “companies don’t have the context anywhere,” Aily hires former business operators from finance, supply chain and R&D rather than relying on generalist consultants. “I tried many times to hire consultants and I have to say that most of them failed,” Anghelina said. “You survive one meeting... with general knowledge. But then the second meeting is bad.” Instead of building an abstract data catalog first, Aily starts from a concrete business decision, such as a real-time forecast, and fixes only the data required for that call. She expects enterprises to repeat the old data-lake mistake with the semantic layer, the map of what corporate data means, by trying to build it in the abstract. “Nobody will build it,” she said. Google’s new Knowledge Catalog is a bet that companies will buy that map from a cloud provider anyway.
Feature Or Position?
Even the strongest independent startups cannot avoid the platforms’ gravity. Aily’s decision intelligence is listed on Google Cloud Marketplace, with its agents running inside Gemini Enterprise. OpenAI names Harvey, Sierra and Decagon among its Frontier partners and lets enterprise customers draw down their OpenAI commitment on 32 partner applications through its new marketplace. The platforms need domain specialists on their shelves, and the specialists need access to where the enterprise budget sits.
Meanwhile, the giants are hiring the human expertise that once protected startups. AWS committed $1 billion to embedding thousands of forward-deployed engineers with customers, and OpenAI dispatches its own forward-deployed engineers into Frontier accounts. Habib contrasts Writer with agent platforms whose engineers “actually do all the work,” betting instead on self-service tools for business teams. The frontier labs are marching the other way, into high-touch services.
That leaves buyers and founders with a clear test: does a startup sell a feature or a position? Basic connectors, cost dashboards and single-session memory are features. They already sit on Google’s and Microsoft’s roadmaps, and selling one means fighting a price war against the company that owns the cloud bill. Portable cross-model memory, multi-cloud neutrality, domain-specific workflow QA and operator-built decision logic are positions. Google cannot credibly promise neutrality about Google, and Anthropic cannot be the neutral custodian of memory a customer wants to take to OpenAI.
Positions erode too. Governance across HR and finance looked like natural ground for an independent startup until Workday made its own permissions and audit trail mandatory for any agent touching its records. Neutrality can also be bought: Opaque pitched confidential, sovereign AI across clouds right up until its team joined Anthropic last week, leaving enterprise buyers to ask what happens to an independent trust layer when one lab absorbs it. And a startup selling through a platform’s marketplace sells on the house’s terms, where OpenAI is not the seller of record and has not disclosed what it takes from each deal.
The evidence from Amsterdam points to a squeeze rather than an extinction. As with the enterprise agents that quietly do real work, the startups that last will not be the ones racing to build a louder general-purpose agent. Inside Google’s new workspace agent, a company can now run a task on Anthropic’s model, behind Google’s firewall, pulling context from a user-keyed memory layer while a Dutch startup’s software audits the output. The giants will keep buying features and shelf space. The survivors will be the companies paid to hold what a platform cannot credibly keep for itself, and paid to doubt the giants in the chain.