Open vs closed AI used to be a single architecture decision that a startup made once and lived with. It is now an ongoing set of choices about which AI models to rent, which to adapt, what to own and how much flexibility to keep as prices and capabilities move every few weeks. That shift is the thread running through four sessions TechCrunch has grouped together for Disrupt 2026, which takes place from 13 to 15 October at Moscone West in San Francisco.
TechCrunch’s own preview puts it simply: “choosing a model isn’t necessarily a one-time architecture decision anymore.” Open models are improving, frontier APIs keep advancing, and “some companies are building products that use multiple models rather than committing to one.” The sessions cover that question at four layers of the stack, from multi-model applications and customised models down to infrastructure and the chips underneath.
This article sets out what each session will tackle, what open vs closed AI really means in 2026, what the market data and list prices say, and a practical decision framework founders can use before, during or instead of the conference. We covered another Disrupt 2026 strand, safety-critical AI with Shield AI, Waabi and GM, last month.
Table of contents
- The Open vs Closed AI Question at Disrupt 2026
- What Open vs Closed AI Actually Means in 2026
- The Market Data Behind the Open vs Closed AI Choice
- The Cost Side of Open vs Closed AI
- Multi-Model Products: Why Open vs Closed AI Is Rarely Either-Or
- Rent, Customise or Build: How Much of the Stack to Own
- How Nvidia Frames the Open vs Closed AI Trade-Off
- When Open vs Closed AI Reaches the Chip
- A Founder’s Open vs Closed AI Decision Framework
- What UK Founders Should Take From the Open vs Closed AI Debate
- Open vs Closed AI FAQ
- References
The Open vs Closed AI Question at Disrupt 2026
TechCrunch says Disrupt 2026 will host more than 200 sessions across six stages, plus roundtables and breakouts, with over 10,000 attendees, 250 speakers and 300 exhibiting startups. Four of those sessions frame the open vs closed AI decision from different angles.
Four sessions, four layers of the stack
| Session | Speakers | Stage | Layer |
|---|---|---|---|
| The Real Tokenmaxxing: How the Best AI Companies Navigate a Multi-Model World | Mo Jomaa (CapitalG), Vipul Ved Prakash (Together AI), Zuzanna Stamirowska (Pathway) | Builders Stage | Applications using several models |
| Which AI Should Your Company Actually Deploy: Rent, Customize, or Build | Manos Koukoumidis (Oumi) | Real World AI Stage | How much of the stack to own |
| Building AI Startups Worth Betting On | Nader Khalil and Sydney Sykes (Nvidia) | Builders Stage | Frontier APIs against open weights |
| When AI Starts Designing Its Own Hardware | Anna Goldie and Azalia Mirhoseini (Ricursive Intelligence) | Disrupt Stage | Chips and infrastructure |
Why the open vs closed AI question changed
Three things turned a one-off choice into a continuous one. Open-weight releases now arrive every few weeks, with permissive licences that let startups download and keep a model. Frontier APIs keep getting cheaper at the same quality tier. And routing software makes it practical to send different jobs to different models inside one product. The open vs closed AI question has stopped being “which camp?” and become “which model, for which job, this quarter?”
What the preview leaves out
TechCrunch’s piece is an events preview, so it raises the questions without answering them. It gives no prices, no adoption figures and no framework of its own beyond Oumi’s promise of “three decision principles”. The rest of this article fills those gaps with published numbers, so founders can arrive at the sessions with their own view.
What Open vs Closed AI Actually Means in 2026
“Open” and “closed” are shorthand for several distinct options, and most of the confusion in the open vs closed AI debate comes from mixing them up.
Closed: frontier models behind an API
A closed model is one you can only use through its maker’s API or apps. You never see the weights, you cannot run it on your own hardware, and the provider can change the price, behaviour or terms. In return you get the strongest models on the market, with no infrastructure to run. OpenAI’s GPT-6 family, Anthropic’s Claude 5.5 family and Google’s Gemini models are all closed in this sense.
Open: weights you can download and keep
An open-weight model publishes its trained weights under a licence that lets you download, run and usually fine-tune it. Licences vary. Aleph Alpha’s Kolibri-1, released this month, uses Apache 2.0; DeepSeek’s V4-Flash uses MIT. Some open-weight licences add use restrictions, so “open” does not always mean “do anything you like”. Our guide to open-weight AI models compares the leading releases and their licences.
The options in between
Between those poles sit two popular middle routes. Hosted open weights, sold by inference companies such as Together AI, give you an open model through an API, often far cheaper than a frontier model, with the option to take the weights elsewhere later. Fine-tuned models, whether open or through a closed provider’s tuning service, let you adapt behaviour to your data. Each route moves the open vs closed AI dial a different distance.
| Route | Control | Up-front effort | Lock-in risk | Typical fit |
|---|---|---|---|---|
| Frontier API | Low | Lowest | High | Hardest reasoning, fastest start |
| Hosted open weights | Medium | Low | Low | High-volume, cost-sensitive jobs |
| Fine-tuned open model | High | Medium | Low | Narrow tasks with proprietary data |
| Self-hosted open model | Highest | High | Lowest | Data residency, offline or regulated work |
| Own model from scratch | Total | Very high | None | Rare; only when the model is the product |
The Market Data Behind the Open vs Closed AI Choice
Two well-known data sets on open vs closed AI seem to point in opposite directions. Reading them together explains a lot about where startups sit.
Enterprises have leaned closed
Menlo Ventures’ 2025 State of Generative AI in the Enterprise report, published in December 2025 and based on about 500 US enterprise decision-makers, estimated that Anthropic held 40% of enterprise LLM API spend, OpenAI 27% and Google 21%. The remaining 12% was spread across Meta’s Llama, Cohere, Mistral and a long tail. Menlo found that open-source share of enterprise use fell from 19% to 11% over the year.
The three closed labs together took 88% (40 plus 27 plus 21). Menlo also found that Chinese open models made up just 1% of total enterprise usage, about a tenth of enterprise open-source use.
Usage of open models is still rising
The other side of the open vs closed AI story comes from usage rather than spend. Dealroom’s summary of Together AI’s $800 million Series C, announced on 1 July 2026, cites McKinsey research that open-source usage tripled over the year to mid-2026, with nearly three-quarters of organisations expecting to use more. The same summary puts Together AI’s annual bookings above $1.15 billion in the second quarter.
Both can be true at once. Spend share counts dollars, and frontier tokens cost many times more than open ones, so a company can run most of its traffic on cheap open models while most of its bill still goes to a closed lab. Founders should read every open vs closed AI statistic with that distinction in mind.
Startups behave differently from enterprises
Menlo noted the growing popularity of Chinese open models among startups, singling out Qwen3 and GLM, even while enterprises stay cautious. It cited Airbnb relying on Qwen for user-facing features and Cursor using Qwen as the base for an internal model. Startups have fewer procurement hurdles and tighter margins, which pushes them towards whatever model is cheapest for each job. In practice, many startups settle open vs closed AI job by job rather than company-wide.
The Cost Side of Open vs Closed AI
Price is the most concrete part of the open vs closed AI decision, and the gap is wider than many founders expect.
List prices per million tokens
The table below combines list prices we have published from each provider’s launch materials. Frontier prices were current at the end of September 2026; hosted open-weight prices are from our August survey and move often, so check them before you budget.
| Model | Type | Input, $ per 1M tokens | Output, $ per 1M tokens |
|---|---|---|---|
| GPT-6 Astra | Closed | 10.00 | 50.00 |
| Claude Opus 5.5 | Closed | 4.00 | 20.00 |
| GPT-6.1 Sol | Closed | 2.00 | 10.00 |
| Claude Sonnet 5.5 | Closed | 2.00 | 10.00 |
| Kimi K3 | Open weights, hosted | 3.00 | 15.00 |
| Mistral Large 3 | Open weights, hosted | 0.50 | 1.50 |
| DeepSeek V4-Pro | Open weights, hosted | 0.435 | 0.87 |
| DeepSeek V4-Flash | Open weights, hosted | 0.14 | 0.28 |
A worked example
Take one job that sends one million input tokens and gets 100,000 output tokens back, with no caching. At list prices, GPT-6 Astra costs $10 plus $5, or $15. Claude Opus 5.5 costs $4 plus $2, or $6. GPT-6.1 Sol and Claude Sonnet 5.5 each cost $2 plus $1, or $3. Kimi K3 costs $3 plus $1.50, or $4.50. Mistral Large 3 costs $0.65, DeepSeek V4-Pro about $0.52 and DeepSeek V4-Flash about $0.17.
On these numbers the most expensive option costs about 90 times the cheapest ($15 divided by $0.168). Even the mid-tier closed models cost about 18 times as much as DeepSeek V4-Flash for this job. Kimi K3 shows that “open” does not automatically mean cheap: a frontier-class open model at its listed API price can cost more than a mid-tier closed one.
Costs the token price leaves out
Token prices flatter open models. Running them well takes engineering time, evaluation suites, monitoring, and sometimes GPUs you pay for whether they are busy or not. Closed APIs bundle all of that. The honest open vs closed AI comparison is total cost per successful task, including the people who keep the system working, not price per token.
Multi-Model Products: Why Open vs Closed AI Is Rarely Either-Or
The first session, “The Real Tokenmaxxing”, starts from the premise that a company “needs to choose one in the first place” is outdated. TechCrunch says it will explore “why companies are using multiple models; how they balance cost, performance, and flexibility; and when open models can outperform proprietary alternatives.”
Routing each job to the right model
Most AI products contain jobs of very different difficulty: classifying a support ticket, summarising a document, planning a multi-step agent task. Sending all of them to a frontier model wastes money; sending all of them to a small model wastes quality. A routing layer that picks the model per request is now standard practice, and it turns open vs closed AI from a company-level bet into a per-request setting.
Together AI’s open-model pitch
Together AI, whose co-founder and chief executive Vipul Ved Prakash is on the panel, sells exactly that middle ground: hosted open models such as DeepSeek, Nemotron and Kimi. Its July round valued it at $8.3 billion after an $800 million raise led by Aramco Ventures. The company says customers report savings of six to 60 times against closed systems, and that Decagon cut inference costs sixfold after switching. Those are vendor figures, but they show the scale of the gap for high-volume work.
The investor and research view
Mo Jomaa represents CapitalG, Alphabet’s independent growth fund, which brings the investor’s question: does a multi-model strategy improve margins enough to change a company’s value? Zuzanna Stamirowska’s Pathway works on live-data AI and published the “Dragon Hatchling” research architecture in 2025, a reminder that new model designs keep arriving and that today’s choice may not be next year’s.
Rent, Customise or Build: How Much of the Stack to Own
Manos Koukoumidis of Oumi will run the most practical of the four sessions, using audience polls and startup scenarios to “evaluate frontier APIs, customized open weights, and owning AI outright”. Attendees are promised “three decision principles”.
Who is asking the question
Oumi launched in January 2025 as a public benefit corporation with a $10 million seed round led by Venrock and Obvious Ventures, backed by researchers from 13 universities including Carnegie Mellon, Stanford and MIT. Koukoumidis previously led Google Cloud’s PaLM work and Gemini’s alignment and safety efforts. Oumi’s platform supports the whole model lifecycle, from data curation and training to evaluation and deployment, so its bias is towards giving companies more ownership.
The three options compared
| Option | Choose it when | Watch out for |
|---|---|---|
| Rent a frontier API | You need the best reasoning now and are still finding product fit | Margin squeeze at scale; terms and models change under you |
| Customise open weights | A narrow task, proprietary data and enough volume to matter | Evaluation and maintenance become your job |
| Build and own outright | The model itself is the product or the moat | Time, talent and compute far beyond most startups |
Customisation as a moat
TechCrunch frames the session around “when customization can become a competitive advantage”. For most startups the answer is narrow tasks where their data is unique: a legal startup’s annotated clauses, a logistics company’s routing history. A fine-tuned open model trained on that data can beat a general frontier model on that one job, at a fraction of the cost, and a competitor cannot rent the same advantage.
How Nvidia Frames the Open vs Closed AI Trade-Off
Nader Khalil, Nvidia’s director of developer technology, and Sydney Sykes, its global head of VC partnerships, will look at “what founders are choosing today, the trade-offs between frontier APIs and open-weight models, and how those choices can affect product strategy and long-term differentiation.”
Nvidia’s stake in openness
Nvidia sells the hardware that both camps run on, but it also publishes open models of its own, including its Nemotron family, and it hosted the July 2026 “Open Weights and American AI Leadership” letter, which carried 235 signatories. Its venture arm, NVentures, invests in startups on both sides. That makes Nvidia one of the few voices in the open vs closed AI debate that profits whichever way founders go, although more open models running on more GPUs suits it well.
Differentiation competitors cannot copy
The session’s sharpest open vs closed AI question is differentiation. If your product is a thin layer on a closed API, a rival can rebuild it on the same API in weeks. Open weights do not fix that on their own, since everyone can download them too. What lasts is the combination of proprietary data, evaluation know-how and workflow design, and the open vs closed AI choice matters mostly for how much of that combination you can control.
The capability gap is narrowing but real
Open models have trailed the frontier by months rather than years. OpenRouter said in June 2026 that open-weight models had shown a consistent three-to-six-month gap behind the frontier for over 18 months. For many products, a model that is six months behind is more than good enough; for products that compete on the hardest reasoning, it is not.
When Open vs Closed AI Reaches the Chip
The fourth session goes below the model. Anna Goldie and Azalia Mirhoseini, founders of Ricursive Intelligence, will discuss “how AI is optimizing chips and hardware” and “what an increasingly open AI ecosystem could mean for the infrastructure underneath it.”
Who Ricursive is
Goldie and Mirhoseini created AlphaChip at Google, which they say has been used across four generations of Google’s TPU and by outside chip companies. Ricursive raised a $35 million seed round at a $750 million valuation in December 2025, led by Sequoia, then a $300 million Series A at a $4 billion valuation in January 2026, led by Lightspeed with NVentures among the investors. It is building a platform meant to let AI keep improving the chips it runs on, which in turn improve the AI.
Why chips matter to model choice
Hardware sets the price floor for every model. If AI-designed chips speed up hardware cycles, inference gets cheaper faster, which helps whoever is running models at scale. Cheaper inference narrows the cost advantage of open models over closed ones at the same quality, but it also makes self-hosting open weights more affordable. Either way, the economics behind the open vs closed AI choice will keep moving.
A Founder's Open vs Closed AI Decision Framework
Disrupt’s sessions will add colour, but founders can make most of the open vs closed AI decision on paper. These questions work for an MVP and for a scaling product.
Five open vs closed AI questions to answer first
First, what does a successful task need: frontier reasoning, or a dependable narrow skill? Second, what volume do you expect in a year, and what does the worked cost example look like at that volume? Third, does any customer, contract or regulation require data to stay in a region or on your own infrastructure? Fourth, which part of your product is hard to copy, and does model choice affect it? Fifth, how quickly could you switch providers if prices or terms changed tomorrow?
Starting points by scenario
| Scenario | Sensible starting point | Revisit when |
|---|---|---|
| Pre-product-fit MVP | Frontier API behind your own interface | Monthly model spend becomes material |
| High-volume simple tasks | Hosted open weights | Quality complaints rise |
| Agentic, multi-step product | Router: frontier for planning, open for steps | A cheaper model passes your evaluations |
| Regulated or sovereign data | Self-hosted open weights | A closed provider offers in-region hosting |
| Unique proprietary data | Fine-tuned open model | Frontier models close the gap on your task |
Design for switching from day one
The cheapest insurance in the open vs closed AI decision is portability. Put every model call behind one internal interface, keep prompts and tools provider-neutral where you can, and build an evaluation set from real tasks so you can test a new model in an afternoon. A startup that can swap models in a day can take advantage of every price cut and every new open release. If you are scoping that architecture for a first product, our MVP development team builds it in from the start.
Read the licence and the terms
Open-weight licences differ, and closed providers’ terms restrict how their outputs may be used, including for training competing systems. Before you fine-tune on one model’s outputs or ship a product built on an open model, read the licence and the provider terms. Our piece on Garry Tan and model distillation shows how quickly that becomes a strategic issue.
What UK Founders Should Take From the Open vs Closed AI Debate
Disrupt is a US event, but the open vs closed AI decision looks slightly different for a UK startup.
Data residency and sovereignty
UK and European customers increasingly ask where data is processed. Self-hosted open weights answer that question directly and keep cybersecurity controls inside your own perimeter. European releases such as Kolibri-1, built for German and English sovereign deployments, are designed for exactly that. Closed providers are adding regional processing, often at a premium; OpenAI, for example, charges 10% more for regional processing on GPT-6.1 Sol.
Budget in pounds, plan in tokens
Most model prices are set in dollars, so currency moves change a UK startup’s costs even when list prices do not. Model your costs per successful task, convert at a cautious rate, and keep the switching option open. For a structured view of where AI fits a business plan, our AI strategy service starts with exactly these trade-offs.
Use the sessions even if you are not there
TechCrunch typically publishes coverage of major Disrupt sessions. Oumi’s three principles and Nvidia’s view of what founders are choosing will be worth reading in full, alongside the multi-model panel, because they come from people who see many startups’ stacks at once.
Open vs Closed AI FAQ
When and where is TechCrunch Disrupt 2026?
From 13 to 15 October 2026 at Moscone West in San Francisco, according to TechCrunch.
Which Disrupt 2026 sessions cover open vs closed AI?
Four: “The Real Tokenmaxxing” with CapitalG, Together AI and Pathway; Oumi’s “Rent, Customize, or Build”; Nvidia’s “Building AI Startups Worth Betting On”; and Ricursive Intelligence’s “When AI Starts Designing Its Own Hardware”.
Are open-weight models cheaper than closed models?
Usually per token, sometimes by a large margin, but not always. A frontier-class open model at its listed API price can cost more than a mid-tier closed model, and running open models yourself adds engineering cost.
Are enterprises moving to open models?
By spend, not yet. Menlo Ventures estimated open-source share of enterprise use fell from 19% to 11% in 2025. By usage, open models are growing fast, especially among startups.
Should a startup use more than one model?
Usually yes, once it has real volume. Routing simple jobs to cheaper models and hard ones to frontier models cuts cost without hurting quality, provided you have evaluations to check it. That is how most teams now handle open vs closed AI in practice.
What is the safest first step?
Put every model call behind one internal interface and build a small evaluation set from real tasks. That keeps the open vs closed AI decision reversible.
References
TechCrunch: Open or closed AI? How founders are choosing what to build on at TechCrunch Disrupt 2026
TechCrunch Disrupt 2026 Agenda
Menlo Ventures: 2025 State of Generative AI in the Enterprise
Dealroom: Together AI raises $800M Series C at $8.3B valuation
Crunchbase News: Ricursive Intelligence Lands $300M Series A at $4B Valuation
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.
