Theory Ventures' Tomasz Tunguz on AI value accrual, blockchain investing, and the moat question in a world of fast-moving foundation models

Summary

Tomasz Tunguz, founder and general partner of Theory Ventures, lays out a clear-eyed view of where AI value is accruing, why moats are harder to establish than they appear, and how blockchain infrastructure is quietly becoming a serious financial asset class.

Theory Ventures: Structure and Scale

Theory raised Fund I at $235 million in late 2022, during the Fed's 500-basis-point rate hiking cycle, then followed with Fund II at approximately $450 million two years later. The firm runs a concentrated model, backing roughly 10 companies per fund at the Series A, with average check sizes scaling from $13–14 million in Fund I to $22–25 million in Fund II. The firm operates more like a company than a traditional partnership, with dedicated teams for internal software, portfolio sales buildout, and a 200-person buyer network.

AI Value Accrual: Feedback Loops Over Raw Data

Proprietary data assets like YouTube or GitHub's code corpus matter early, but the durable advantage is the feedback loop. The signal that matters is what users actually do with a prompt and whether the model output was correct. That telemetry compounds over time and is what separates leading models as the field matures. DeepSeek's ability to distill outputs by calling competitor APIs at scale is cited as evidence that raw training data is less defensible than previously assumed. Tunguz flags OpenAI's decision to release a Deep Research API as surprising given distillation risk, particularly after the DeepSeek episode.

On Google specifically, Veo 3's YouTube data advantage is real but undermined by capacity constraints. Both Google and Microsoft have acknowledged in public earnings that data center buildout is the binding constraint, not model quality. Google is responding by subsidizing Gemini fine-tuning and offering a more generous free tier on its command-line interface, a deliberate attempt to maximize usage telemetry. A reported arrangement where OpenAI picked up Coreweave compute capacity that Microsoft relinquished, routed through Google infrastructure, suggests at least some OpenAI training workloads are now running on Google hardware.

Reinforcement Learning as Infrastructure

Reinforcement fine-tuning and test-time compute are framed as foundational model-layer stories, not startup opportunities. The practical implication is a tiered compute economy where reasoning budget scales with task complexity and willingness to pay. This is already implicit in how users choose between fast, cheap responses and slower deep research modes under flat subscription pricing like OpenAI's $200 per month plan. The next evolution prices that compute explicitly, with users paying per query based on depth and speed required.

The Moat Problem

Memory is emerging as a meaningful consumer moat. The more time a user spends with a given model, the more the switching cost compounds. In enterprise, the question is whether companies will consolidate memory within a single foundation model provider or insist on portability across clouds and models. That structural choice will determine how enterprise AI software gets built. Tunguz also highlights the erosion of legacy switching-cost moats: agentic workflows capable of migrating payroll or CRM data in a 24-hour window could hollow out the operational friction that has protected incumbents like Salesforce and Workday.

The broader enterprise context layer, sitting between the model and the UI, is identified as an underappreciated battleground. Whoever controls how enterprise data is structured and accessed will have significant leverage over AI product performance.

Startup Metrics Under Pressure

The speed at which AI companies are scaling is breaking traditional underwriting frameworks. A company that goes from zero to $100 million ARR in 12 months may be valued at $400–500 million with only one or two quarters of retention data available. Ten years ago, a company reaching that valuation would typically have had 24 months of longitudinal net dollar retention to analyze. Top-quartile NDR historically ran at 125% post-2022; that benchmark still matters but is nearly impossible to validate at Series A given current growth velocities. Tunguz argues investors are effectively taking seed-level risk at Series A prices for the fastest-growing AI companies.

The credit dynamic compounds the distortion. Hyperscaler investments structured as partial cloud credits, where, for example, a $1 billion commitment includes $600 million in cloud credits, create circular revenue flows that echo dot-com era dynamics. Startups spending those credits with the same hyperscaler that invested in them inflates reported consumption metrics and can obscure whether real third-party demand exists.

Blockchain: Infrastructure, Stablecoins, and Compliance

Theory was the second-largest investor in SUI, now valued at $40–50 billion. The firm also holds a position in Alium, which supplies blockchain data to Stripe and three of the top five Bitcoin ETFs. Tunguz points to Ethereum's Q1 performance last year, generating $400 million in free cash flow, making it the largest free cash flow producer on a percentage basis among publicly traded software companies and sixth largest in absolute terms, as evidence that blockchain database infrastructure can generate software-scale economics.

Stablecoins are flagged as the dominant theme of the past 18 months, with the sector now the 15th largest buyer of U.S. Treasuries. Longer term, Theory's thesis is that within 10 years every major software product will carry a Web3 component, driven by the complexity of multi-jurisdictional data compliance. Customer-custodied data with selective computational access grants would allow software vendors to sidestep the compliance burden imposed by Germany's data locality rules, UAE regulations, and 26 U.S. state-level frameworks.

Nvidia and the ASIC Question

Nvidia's PE multiple has spiked three times in its history, around gaming, blockchain, and now AI, with each cycle compressing back toward baseline. The classic hard-disk analogy, sell into the ramp before excess supply hits, has not played out this cycle, partly because demand has continued to outpace supply. Nvidia's push into managed cloud via DGX Cloud is read as a necessary diversification toward a longer-term business model. ASICs are viewed as a legitimate and growing part of the compute stack, with Google's TPUs, Amazon's Trainium and Inferentia, and Microsoft's custom silicon all indicating that hyperscalers are committed to reducing Nvidia dependency. The key architectural distinction noted is that AMD chips have historically been optimized for mid-size models while H100s remain better suited to 100 billion-plus parameter workloads due to differences in memory architecture.