Semafor14 min
ICONIQ Founding Partner Divesh Makan at Silicon Valley & The World | Semafor
Interview, Fireside Chat
Anthropic Investment Lessons
- Iconic incorrectly assumed the cloud infrastructure layer (Azure/AWS) would dominate value creation rather than the frontier AI model owners.
- Investors underestimated the sustainability and velocity of scaling laws, which have proven far more effective within their investment timeframe than anticipated.
- Iconic initially rejected Anthropic multiple times, including three months prior to their $10B valuation round when the price was one-third of the final investment.
- The decision to invest shifted after recognizing a structural change in enterprise adoption, specifically the emergence of first-party direct product usage.
- The firm acknowledges the cultural difficulty of reversing a "no" decision without defensiveness, citing it as a key lesson for their investment culture.
AI Value Chain and Application Strategy
- Iconic identifies a four-tier AI value hierarchy: Energy, Chips, LLMs/Processing, and Applications.
- Current thesis posits that significant value will migrate to the application layer as the ecosystem matures, similar to the shift from hardware to apps in the smartphone era.
- Iconic invests in specific application-layer companies, including Eleven Labs (founded by Mati) and Open Evidence, noting that initial underinvestment in "show and tell" phases required strategic correction.
- Future investment strategies must determine the economic boundaries between LLM providers and application developers.
Capital Markets: Connected vs. Circular Financing
- The speaker rejects the "circular financing" narrative surrounding NVIDIA and AI infrastructure, classifying current spending as "connected financing" similar to historical railroad and fiber optic builds.
- The framework assumes a conservative 10-year horizon for scaling laws to continue, justifying current capital expenditure pipelines.
- Hyperscalers (the "five large" companies) are funding over $1 trillion in compute spend this year primarily via operating cash flow rather than debt.
- NVIDIA's pricing power remains intact, evidenced by a recent 15% price increase on GPUs that did not deter purchasing volumes.
- Total projected compute infrastructure spend is estimated at $3 trillion over the coming years, driven by the demand for chips, memory, and optical networks from suppliers like TSMC and ASML.
Geopolitics and US-China AI Dynamics
- US regulatory scrutiny on Chinese AI exports is expected to intensify with sharper definitions of allowable technology transfers.
- Iconic estimates China's LLM capabilities are currently 12 months behind Western frontier labs (Anthropic/OpenAI).
- Cultural Divergence in R&D:
- US labs engage in "pontification" or open-ended exploration without immediate utility, a practice permitted by the US market structure.
- Chinese entities operate under strict utility requirements; spending money without a defined purpose risks corporate or government repercussions.
- Strategic Adaptation to Scarcity:
- Chinese labs, facing compute constraints, are adopting efficiency techniques such as model distillation that are increasingly being adopted by US firms.
- Chinese AI spending claims are structurally distinct; a major tech executive cited a personal engineer salary cap of $15M/year, contrasting with US billion-dollar compensation packages.
- Security Assessment:
- The US may be over-paranistic regarding China surpassing it technologically in generative AI.
- Conversely, the US is likely under-paranistic regarding the risk of Chinese AI models being used for infiltration and domestic disruption.
- The Chinese government prioritizes party control; while AI aids information gathering, it is viewed as a threat to regime stability due to its potential to educate the populace and generate dissent.