The Twenty Minute VCSam Altman, Arthur Mensch and more discuss:Which Startups Are Threatened vs Enabled by OpenAI?|E1156
CHAPTERS
- 0:00 – 1:03
Why foundation models will consolidate—and differentiation shifts to personalization
Sam Altman compares today’s model boom to the early auto industry: lots of competing players before consolidation. He argues base models will become a small set of expensive, complex providers, while enduring advantage moves to deeply personalized, integrated assistants.
- •Early-stage markets have many entrants before a shakeout
- •Model training at scale will be concentrated among a small number of providers
- •Competition should drive models to get better, cheaper, and faster
- •Long-term differentiation is personalization + life context + integration, not the base model
- •Near-term priority: keep improving the base model on a steep capability curve
- 1:03 – 2:20
Mistral’s view: two opposing forces thinning both the model and app layers
Arthur Mensch describes a tension: better models make it easier to build vertical apps, but cheaper models compress pricing power at the model layer. Mistral’s strategy is to build a platform on top of strong models to enable many vertical applications.
- •Rising model capability lowers the difficulty of building verticalized apps
- •Efficiency gains compress “dollars per intelligence unit”
- •These trends can thin both application differentiation and model margins
- •Mistral bets the model layer remains large enough to matter
- •Platform-building is positioned as the enabler for vertical app ecosystems
- 2:20 – 3:59
Investor anxiety: foundation models as rapidly depreciating ‘power stations’
Tom Hulme argues model tech is commoditizing so quickly that the economics resemble building a power station that depreciates in months. He highlights the limited edge among teams using similar GPUs and points to Meta’s scale and open-sourcing as a major accelerator of commoditization.
- •Training is capex-heavy; inference is the ‘power’ output
- •Competitive parity: similar hardware (GPUs) and marginal gains
- •Asset depreciation happens over weeks/months, undermining fundamentals-based investing
- •Meta’s H100 scale + open source (Llama) intensifies price/capability pressure
- •GV prefers infra/picks-and-shovels and apps over foundation model bets
- 3:59 – 5:37
Can you still make money in foundation models? Momentum vs fundamentals
Hulme distinguishes between making money via market momentum (liquidity and markups) versus durable fundamentals in a fast-commoditizing layer. He frames GenAI largely as a sustaining innovation that spreads across industries rather than causing internet-like creative destruction.
- •OpenAI-type positions have offered strong secondary liquidity and markups
- •Momentum investing may work if a leader stays ahead of the pack
- •Fundamentals are challenged by rapid commoditization
- •Clay Christensen lens: GenAI as sustaining innovation
- •Impact seen more in cost reduction and incremental product improvements than industry reset
- 5:37 – 6:18
End-state thesis: models become utilities owned/distributed by cloud giants
Harry proposes a future where cloud providers become the cash cows, acquire model companies, and bundle/give away models to drive compute consumption. Hulme agrees: models look like utilities and clouds will monetize by hosting and charging for usage on their compute stacks.
- •Cloud providers may buy/acquihire foundation model companies
- •Bundling models could drive demand for underlying cloud compute
- •Utility-like dynamics: standardized service, competitive pricing pressure
- •Clouds already charge for model access across their platforms
- •Distribution + compute ownership becomes central power
- 6:18 – 7:12
Operator perspective (Intercom): value flows to infra, but portability matters
Des Traynor notes that today a lot of value is captured by infrastructure/model providers (e.g., OpenAI) as application companies pay upstream. He emphasizes that LLMs aren’t yet equal, so the ability to switch models quickly is strategically important—but winning takes more than being model-agnostic.
- •A meaningful share of economics currently accrues to model/infra providers
- •Intercom stress-tests multiple LLMs; performance differences remain
- •Strategic need: fast migration when a better model emerges
- •Model-agnosticism alone isn’t sufficient to win
- •Expectation of major moves from platform players (Amazon, Apple); skepticism about some releases being reactive
- 7:12 – 9:12
Would you invest in OpenAI at $90B? Concerns about cloud bundling and moats
Traynor and Hulme both hesitate on investing at a $90B valuation, largely due to commoditization risk and cloud-provider distribution advantages. Hulme outlines what could make a foundation model defensible: real memory, durable consumer stickiness, or meaningful agentic capability beyond ‘more compute.’
- •Amazon risk: acquire a top model and bundle into EC2 for easy enterprise adoption
- •Cloud-native privacy and distribution can leapfrog adoption concerns
- •Meta + cloud arms race make advantages feel ephemeral
- •Potential moats: persistent memory/personalization and agentic ‘doing’ capabilities
- •If it’s just scaling compute like everyone else, ROI is hard to justify
- 9:12 – 10:06
Learning from cloud history: infrastructure vs applications value capture (Tunguz)
Tomasz Tunguz analyzes Web 2.0 outcomes: the top three cloud infrastructure businesses and the top 100 cloud apps ended up with similar total market cap. For investors, the application layer offers more shots on goal because it contains many winners rather than a few concentrated incumbents.
- •Top 3 cloud infra businesses total market cap roughly matches top 100 cloud apps
- •Infrastructure is concentrated; applications are diverse
- •Equal value pools can exist at both layers, but winner distribution differs
- •Investor odds improve with breadth and varied customer needs at the app layer
- •Implication: application-layer investing may offer better portfolio-level outcomes
- 10:06 – 10:54
Emad Mostaque’s forecast: only ~5–6 model trainers survive; capital intensity decides
Emad predicts a small group of foundation model companies will dominate within a few years, mostly tied to the largest tech platforms. He questions how independents can keep up against players like Google with massive annual AI spend and talent budgets.
- •Prediction: only a handful of foundation model companies will remain
- •Likely winners include big platforms plus key ecosystem players (e.g., NVIDIA)
- •Anthropic praised technically, but business-model pressure highlighted
- •Google’s scale (AI spend, DeepMind payroll) creates daunting competitive asymmetry
- •Survival depends on sustained access to capital, compute, and distribution
- 10:54 – 11:51
Two startup playbooks: build for static models vs ride the improvement curve
Sam Altman and Brad Lightcap outline two strategies: assume models won’t get much better and build lots of scaffolding, or assume rapid improvements continue and design to benefit from them. Lightcap warns that startups built on the ‘models won’t improve’ assumption risk being overtaken as the base models advance.
- •Strategy A: assume model stagnation and build many compensating layers
- •Strategy B: assume models keep improving rapidly and architect accordingly
- •OpenAI’s core progress can ‘steamroll’ startups reliant on model limitations
- •Most companies should rationally bet on continuing model improvement
- •Founders must align product design with the expected capability trajectory
- 11:51 – 12:28
A practical ‘steamroll’ test: are you excited about 100× better models?
Lightcap proposes a simple diagnostic: if a company is thrilled by massive model improvements, it’s likely positioned to benefit rather than be displaced. Companies that actively demand early access to new models often have a clearer path to compounding advantage with better intelligence.
- •Key question: would 100× better intelligence materially help your product?
- •Signal: companies eager for next models can articulate clear upside
- •If improvements remove your core differentiation, risk is high
- •Founders should be able to explain how better models accelerate their roadmap
- •Model progress is predictable enough to stress-test business defensibility
- 12:28 – 13:22
Thin wrapper vs thick wrapper: solve an end-to-end vertical problem OpenAI won’t
Des Traynor argues thin wrappers—filling temporary platform gaps—are like picking up coins on train tracks: eventually the platform catches up. Thick wrappers win by solving the full workflow end-to-end in a domain where OpenAI won’t invest deep integration effort (e.g., regulated or integration-heavy verticals).
- •Thin wrappers exploit temporary gaps and are structurally fragile
- •Platforms reach ‘good enough for everyone’ and close feature gaps over time
- •Thick wrappers own the workflow and deliver an end-to-end solution
- •Choose domains requiring deep integrations and operational depth
- •Differentiate by nailing the use case, not building demos
- 13:22 – 14:13
Where enduring value accrues: own the end user and compound application leverage
Sarah Tavel argues most value will be created and captured in the application layer because user ownership enables compounding value delivery over time. While model competition may form an oligopoly, she focuses on apps as the primary locus of durable value capture.
- •Owning the end user enables ongoing value expansion and monetization
- •Model layer faces intense competition and uncertain structure (oligopoly vs more)
- •Applications can build durable relationships and distribution
- •Value capture follows user ownership more than underlying commodity capability
- •Investor focus: app-layer businesses that compound over time
- 14:13 – 18:19
Beyond ‘copilots’: incumbents’ advantage vs startups’ disruption via outcomes
Tom Blomfield says defensible AI startups are mostly traditional software plus AI, deeply embedded in industry workflows, tooling, and regulation—areas OpenAI won’t customize heavily. A Thrive guest argues copilots are an incumbent strategy (distribution, data, UX), while Tavel adds the disruptive startup move is shifting pricing to selling outcomes—‘doing the work’—not per-seat software.
- •YC view: defensibility comes from deep industry embedding and workflow integration
- •Most ‘AI apps’ are 80–90% classic software, 10% AI, plus domain/regulatory depth
- •Copilots fit incumbent advantages: distribution, data, UX control, existing business models
- •Startups may need orthogonal strategies rather than ‘copilot for X’ positioning
- •AI enables outcome-based pricing: selling completed work products vs per-seat productivity tools
