AI startups can compete with OpenAI and other large platforms by specializing in portable memory, provider-neutral governance, domain-specific workflow checks or data organized around a specific business decision. In interviews at HumanX Amsterdam reported by Forbes, founders and executives advanced these as possible strategies, not proven routes to lasting advantage.
Four possible ways to compete
| Position | Example | Potential role | Competitive pressure or limitation |
| Portable memory | Walrus Memory from Mysten Labs | Shares encrypted memory among agents granted delegate keys. | Exporting memory does not ensure another model will interpret the context the same way. |
| Provider-neutral governance | Deeploy | Monitors AI systems across an organization; the company says it can run on customer infrastructure. | Governance tools overlap with platform offerings, and Mysten Labs announced an audit-layer collaboration with Google Cloud. |
| Domain-specific workflow checks | Writer’s Enterprise Brain | Encodes shared team memory and compliance rules into regulated workflows. | Google’s legal-workflow preview involved Harvey and Onit, creating overlap between specialist and platform-led products. |
| Decision-first data | Aily Labs | Starts with a business decision, then addresses the data needed to make it. | Google’s Knowledge Catalog is a platform-led alternative. |
Portable memory has an interpretation problem
Walrus Memory is built on Walrus and anchored to Sui. Its design lets users grant different AI agents delegate keys to shared encrypted memory. That could make a change of model less disruptive when a team wants to carry context with it.
Moving the data is only part of that task. A different model may interpret the same stored context differently, so portability alone does not guarantee continuity in how an agent responds. The design centers on user-granted keys; the practical question for a business is whether the receiving model can make useful sense of what those keys let it access.
Governance depends on who can monitor the system
Deeploy executive Roeland Pater described a layer that can sit across an organization’s systems and run on customer infrastructure. In that approach, buyers monitor AI outputs and retain policy decisions. The distinction matters when a company wants oversight across several systems rather than relying on each platform to govern its own agents.
That position does not belong to startups alone. On October 6, 2026, Mysten Labs and Google Cloud announced a collaboration to build Verifiable Agent Arbiter, an evidence layer intended to separate audit logs from one provider. The partnership puts a startup and a major cloud provider on the same project.
Workflow expertise lives in the checks
Writer CEO May Habib argued that domain-specific rules and quality assurance are central to enterprise AI workflows. Writer’s Enterprise Brain encodes shared team memory and compliance rules into regulated processes. That work also involves less glamorous failure points: a connector can break or an authentication token can expire, leaving an automated process failing silently.
For a business, the appeal is a workflow built around its own rules and checks, rather than a general-purpose model alone. Google’s legal-workflow preview, which involved Harvey and Onit, shows how platform companies can also work with specialists in a specific domain.
Start with a decision, not a data catalog
Aily Labs co-founder and CEO Bianca Anghelina described a decision-first approach: begin with a concrete business question, such as a real-time forecast, then address the data needed to answer it. That can keep data preparation focused on a defined use instead of turning into an open-ended effort to organize everything first. Google’s Knowledge Catalog represents a platform-led approach in the same broad area.
What Mysten’s Walrus Memory figures measure
Mysten Labs reported more than 5,000 agents across 4,800 owners since Walrus Memory launched in the second quarter of 2026. By October 20, Mysten reported more than 209,000 stored memories and 4.5 million successful retrievals. The company also reported that 72.9% of owners whose first memory write occurred on or before September 19 retrieved at least one memory 30 or more days later.
Forbes reported that the totals were inferred from encrypted read-write activity and were not independently audited. The figures therefore describe reported activity, not a count of paying enterprise customers. The 72.9% measure applies to the specified first-write cohort and retrieval period.
Platform overlap and startup independence
Several large companies are building in adjacent areas. Forbes reported that Workday’s Trusted Agent Platform provides agents access to HR and finance records. OpenAI’s marketplace included 32 partner applications, with Harvey, Sierra and Decagon named among its Frontier partners. These channels can connect specialist products with enterprise buyers, while also making their reach dependent on a platform’s distribution and terms.
Consolidation is another pressure. Forbes, citing Crunchbase, reported 195 acquisitions of AI startups by venture-backed AI companies through September 29, 2026, more than in all of 2025. It reported 10 acquisitions by OpenAI and five by Anthropic in 2026 through its reporting period. Separately, Opaque announced a license-and-hire arrangement under which its team would join Anthropic. Such transactions can bring specialist teams into larger companies, even as independent startups continue to pursue focused roles.
Distribution remains a separate hurdle from building a useful product. Mysten Labs co-founder Kostas Chalkias raised that issue while discussing competition with established companies: a startup needs a way to reach customers as well as a capability worth adopting.
