Bucky Moore on Lightspeed's $9B in new funds, inference as the next cloud, and why the SaaS apocalypse is overblown

Summary

Lightspeed closes $9B, bets on inference as the next cloud

Lightspeed Venture Partners has closed $9 billion in new funds across venture and growth, and partner Bucky Moore is deploying much of his attention at the early-stage AI layer — both infrastructure and applications. His central argument is that the SaaS apocalypse narrative is overstated, the inference market will dwarf pre-ChatGPT cloud computing, and the defensible application businesses of this era will look less like traditional SaaS and more like Palantir.

The inference thesis

Moore believes inference will become a larger market than cloud computing was before ChatGPT, with "hundreds and hundreds of billions" in annual spend. He expects the inference market to fragment into specialized platforms — one for real-time video models, one for open-source and custom language models, one built specifically for long-running agents. His reasoning is that as agentic workloads grow, the latency constraint driving today's hardware decisions starts to relax. When an agent can run for days rather than milliseconds, throughput matters more than speed, which makes Intel and AMD chips suddenly competitive with Nvidia's Blackwells — and considerably cheaper. More long-horizon agentic workloads means a more heterogeneous hardware ecosystem, and Moore sees that shift as underappreciated.

Power as a bottleneck

Moore flags power as still underrated despite the attention it receives. His specific observation is that neo clouds are being blocked not just by physical capacity but by their ability to raise the credit needed to bring that capacity online — and the quality of their counterparties determines whether debt providers will fund them aggressively. OpenAI as a counterparty unlocks capital easily. A neo lab without near-term revenue does not. Lightspeed has been investing in nuclear through Base Power and expects to stay active there.

Neo labs: the bar is higher

Lightspeed is already an investor in Mistral, Safe Superintelligence, and Thinking Machines, but Moore is candid that the frontier bar has risen sharply. He estimates only three or four labs can genuinely compete at scale. He leaves a door open for exceptions — noting that OpenAI exploited Google's bureaucracy by scaling the transformer before Google could, and that the same dynamic could recur as the large labs get distracted protecting their revenue-generating products. But the second financing round for any new lab, he says, will require both demonstrable commercial progress and genuinely novel research, and absent both, pulling that round together has become much harder.

Application layer: the Palantir model

Moore's framework for backing application-layer companies draws directly on the Databricks/Snowflake-on-AWS analogy. He argues the large model providers will be "perfectly happy" letting roughly 75% of industries sit on top of them — the same way AWS let Snowflake and Databricks build large businesses while consuming AWS infrastructure. The 25% of industries where the labs will compete directly are the ones to avoid. The defensible businesses are in the messy, cross-functional deployments where context engineering is genuinely hard, forward-deployed engineering is required, and customer data flows only to the vendor serving them. That data asymmetry, Moore argues, compounds into a feedback loop that outsiders can't replicate — the same dynamic that made Palantir durable.

On Cursor and the software engineering vertical, Moore frames the open question clearly: whether post-training open-weight models combined with unique user feedback from being an application provider is defensible enough. The challenge is that Cursor still routes its most complex queries to GPT and Claude, creating a margin problem when OpenAI and Anthropic are also shipping competing products. He expects Harvey and Legoura in legal to face the same dynamic. His conclusion is not that these companies fail, but that any vertical reaching the maturation point of AI adoption will hit this tension.

MCP as a new monetization layer

Moore highlights Lightspeed portfolio company Granola as an early example of a new business model: metering MCP server consumption. Every time Claude or another model calls Granola's MCP server, Granola earns a small payment. He describes this as a durable model and expects it to become common as more software companies expose their capabilities through MCP.

Suno and the audio exception

On Suno, another Lightspeed portfolio company, Moore admits he was initially skeptical that AI-generated music would be listenable and says he's been proven wrong. His structural point is that audio models simply don't require the compute scale that language models do, which compresses the lab-scale advantage. That makes audio a category where a focused team can build something the frontier labs don't bother competing on.

Seed economics today

On early-stage pricing, Moore describes the current reality plainly: in the most competitive seed rounds, investors are getting around 10% ownership. Series A valuations are running between $150M and $300M post-money, with little derisking between rounds given how fast companies are moving. Multistage firms like Lightspeed are playing the seed for 10% and then co-leading or leading the Series A to land in the low-to-mid teens. Moore says founders and their valuations dictate terms, not investors — a dynamic he treats as structural, not temporary.

Ghost ships and founder bribes

On the creative acquisition structures circulating in the market — licensing deals, team hires, and dividends that substitute for traditional M&A — Moore notes these are partly a regulatory workaround and partly a product of bankers actively pitching the structure as more executable than a full acquisition. He uses the term "founder bribe" for the pattern of structuring proceeds to disproportionately reward founders and core teams at the expense of investors, and frames it as rational from the acquirer's perspective. He says the jury is still out on whether acquirers actually get the return they're looking for from these structures, but confirms they are becoming more common.

Geopolitical overhang

Moore declines to predict the LP impact of the Middle East conflict but acknowledges Gulf sovereign wealth funds are active Lightspeed LPs and that their priorities could be shifting. He says he has not yet seen the market signal reduced appetite for companies that are genuinely scaling, but treats the risk as real and open-ended. The practical constraint, as he frames it, is that sovereign capital is zero-sum — dollars directed toward oil and defense cannot flow into venture.

Lightspeed's current posture is clear: inference infrastructure and power are the two most structurally underappreciated bets, the frontier model race is largely closed to new entrants without extraordinary research and commercial credibility, and the application-layer opportunity is real but narrow — concentrated in industries that are large, messy, and far enough from the labs' core commercial interests that they won't be adequately served for years.