For AI startups, model selection is becoming less like a permanent bet and more like an operating strategy. Open models are getting better, frontier APIs continue to improve, and some companies are designing products that can use more than one model depending on the task.
That shift gives founders more room to maneuver. It also creates harder questions about cost, performance, infrastructure, ownership, and how much flexibility a startup should preserve as the technology keeps changing.
One model may not cover every job
The old question of open versus closed AI assumes that a startup must choose one side. Increasingly, the more practical question is whether one model is enough for the product at all.
At TechCrunch Disrupt 2026, the session “The Real Tokenmaxxing: How the Best AI Companies Navigate a Multi-Model World” will focus on that issue. Mo Jomaa, partner at CapitalG; Vipul Ved Prakash, co-founder and CEO of Together AI; and Zuzanna Stamirowska, CEO and co-founder of Pathway, will discuss the topic on the Builder’s Stage.
Their conversation will cover why companies are using multiple models, how they weigh cost against performance and flexibility, and where open models can outperform proprietary alternatives.
For founders, the consequences go beyond technical benchmarks. A multi-model approach can affect operating costs, product design, and the speed at which a company can adopt better models when they appear.
The ownership question is getting sharper
Model choice also raises a deeper business question: how much of the AI stack should a company control directly?
Manos Koukoumidis, CEO and co-founder of Oumi, will address that question on the Real World AI Stage in “Which AI Should Your Company Actually Deploy: Rent, Customize, or Build.” The session is built around whether startups should build their own models, when customization can become a competitive advantage, and how founders should evaluate open and proprietary AI.
The source describes a practical format that includes audience polls, startup scenarios, and a framework for comparing frontier APIs, customized open weights, and owning AI outright. Attendees are expected to leave with three decision principles they can bring into architecture discussions.
The trade-off is clear. Building more of the AI stack can give a startup more control and room for differentiation. But relying less on existing models can also demand more time, talent, and resources.
Open and proprietary AI shape product strategy
For some startups, the decision still centers on the core trade-offs between open-weight models and frontier APIs. Those choices can influence infrastructure requirements, cost structure, and the amount of control a company has over the product it is building.
Nader Khalil, Director of Developer Tech at Nvidia, and Sydney Sykes, Global Head of VC Partnerships at Nvidia, will speak on the Builders Stage in “Building AI Startups Worth Betting On.” Their session will examine what founders are choosing today, the trade-offs between frontier APIs and open-weight models, and how those decisions can affect product strategy and long-term differentiation.
That last point matters because differentiation is not only a technical issue. If competitors can easily reproduce a startup’s model setup, the company may need to find defensibility elsewhere. If the model strategy is more tailored, the startup may have more room to create product advantages that are harder to copy.
- Frontier APIs can help teams use advanced capabilities without owning the full model stack.
- Customized open weights can give companies more control over specific workloads.
- Owning AI outright can create more control, but it can also require more resources.
- Multi-model products can let teams match different models to different jobs.
Hardware is becoming part of the AI architecture debate
The model layer is not the only part of the stack under pressure. AI performance depends on the hardware beneath it, and advances in AI are beginning to affect how that hardware is designed.
Anna Goldie, founder and CEO of Ricursive Intelligence, and Azalia Mirhoseini, founder and CTO, will discuss this on the Disrupt Stage in “When AI Starts Designing Its Own Hardware.” Their session will explore how AI is optimizing chips and hardware, why model architecture and hardware are becoming more closely connected, and what an increasingly open AI ecosystem could mean for infrastructure.
For founders, this connection could influence how quickly new AI capabilities become available. Faster chip development could also change the infrastructure choices open to startups building future AI products.
Flexibility may be the central lesson
TechCrunch Disrupt 2026 is taking place October 13-15 at Moscone West in San Francisco. The event includes 200+ sessions across six industry stages, roundtables, and breakouts. More than 10,000 founders, investors, operators, and tech leaders are expected, along with 250+ speakers and 300+ exhibiting startups.
The AI sessions sit inside a broader event built around startup decisions, networking, matchmaking, and dealmaking. But across the AI programming, one theme stands out: founders may need to preserve optionality rather than treat today’s model choice as final.
A startup might use a frontier API now, customize an open model later, or move workloads across several models as the product evolves. The important decision may not be choosing open or closed AI once. It may be building in enough flexibility to keep choosing as the market, models, and infrastructure change.