Descriptions:
Greylock partner sits down with Bloomberg Tech to discuss the firm’s eighteenth fund — a $1.5 billion vehicle dedicated exclusively to early-stage AI entrepreneurs. The interview spans Greylock’s history of AI bets (Anthropic, OpenAI, Base Ten, Crest AI) and lays out the firm’s current investment thesis across three layers: foundation models, AI infrastructure (which the partner expects will be entirely rebuilt around agents), and applications. He points to portfolio company Resolve AI as evidence that agentic systems are already delivering measurable value, with on-call engineering agents at companies like Coinbase and DoorDash autonomously handling infrastructure incidents.
The conversation pivots to the Kimi K3 release and the market reaction it triggered. The partner pushes back against the prevailing narrative in measured terms: he notes that Kimi K3 is actually more expensive on a per-token basis than its predecessor Kimi K2, and that for real-world task completion it uses significantly more tokens than comparable OpenAI or Anthropic models — making cost-based conclusions premature. He draws a parallel to the market’s reaction to DeepSeek earlier in the year, arguing that benchmarks are an imperfect proxy for real-world performance until models have time to percolate through actual use.
The partner also contextualizes the broader open-source moment, noting that Thinking Machines released a model the day before Kimi K3, and framing open-weight releases as net positives for the overall AI economy rather than existential threats to frontier labs.
📺 Source: Bloomberg Tech · Published July 20, 2026
🏷️ Format: Interview
