Value lands where the outcome lives, not where the tool is sold Prathik Naidu of Long Lake opens with the vendor's arithmetic problem. A one-million-dollar AI contract that lifts a client's EBITDA by five million a year creates roughly fifty million dollars of enterprise value at a 10x multiple — and every dollar of it accrues to the client, not the vendor. Selling AI into an enterprise from the outside means clearing security review, coordinating across the organization, integrating with twenty-year-old systems, and then somehow driving adoption and settling on pricing, seat models, and ROI measurement. As he puts it: you're building on the outside and that makes deploying AI really hard. The deeper barrier is that org charts understate how businesses actually run. Vendors integrate with systems of record — CRMs, ticketing tools — but complex businesses run on human judgment: exceptions, edge cases, and tacit knowledge that never made it into any database. A vendor selling software touches the edges of that judgment; the client's P&L absorbs the gains. Oliver Kennon of SemiAnalysis arrives at the same conclusion from the compute side. GPU rental prices look like a cost of doing AI business, but they measure input, not output: GPU hours are the input, intelligence is the output. The rental bill is revenue to the GPU provider and only one input cost to the inference producer, whose real cost also includes serving software, staff, and operations. In both talks, the same law holds: whoever owns and operates the outcome captures the value. Supplying the tool — chips or agents — leaves you upstream of where the money settles. Buy versus build: what trust and distribution cost Building an AI-native firm from scratch is the purest version of the ownership thesis. You improve whole workflows, sell the work itself rather than a tool, and own the outcomes and the P&L. The costs are equally structural: no customers, no distribution, no brand reputation, and real difficulty hiring niche industry expertise. Naidu's example is a health system planning a surgery building — it will pick the trusted firm of twenty years over an AI architecture startup. Trust is not a feature you ship; it is a balance-sheet asset you buy or accrue. Acquisition converts that asset into day-one advantages: veteran operators, live operational data for improving agents, incentive alignment, embedded trust, and scale. Naidu's framing of why ownership changes behavior is blunt: Ownership solves for this. You only win when your team and your business wins. Long Lake is executing this at conviction scale — $7.5 billion raised from Elad Gil, General Catalyst, Alpha Wave, D1 and others, roughly 40 partner businesses across residential services, architecture, HR and more, and a $6.3 billion Amex GBT take-private into corporate travel. The tradeoff deserves honesty. Building gives you greenfield design and full control but you pay for trust and distribution over years; buying pays for them upfront in capital and integration risk. Neither path works without the operational layer — Long Lake runs dedicated M&A, building, and deployment teams whose job is adoption and real business impact, not just time savings. Services economics, software margins The prize for getting this right is enormous because of where services sit economically. Services is a $70 trillion global market, and owning outcomes end-to-end is far more valuable than selling tools to the people doing the work. The reason is margin structure: a software business keeps about $80 of every $100 of incremental revenue; a service business keeps about $20. Naidu calls this a pretty huge, you know, 80% marginal tax on growth — every dollar of service growth drags eighty cents of cost behind it. AI attacks exactly that link. As Naidu argues, AI actually reduces the cost of service while also improving the quality of your service — which breaks the historical coupling between revenue and labor cost and lets service economics start to look like software. Kennon's compute analysis is the same story in a different industry. The eventual North Star metric for intelligence production is price per successfully completed task at required quality, with price per token as today's practical proxy — an output measure, not the GPU-hour input. His B200 example makes operations concrete: two serving stacks on the same chip, same interactivity targets, deliver roughly 200 versus 300 output tokens per second in the P90 scenario. Same chip, same target interactivity, about 50% more output. That's a lot. At the same hourly bill, 1.5x output proportionally lowers spend per acceptable token. Operations are the multiplier on capital in both domains. Cost position is relative — and the clock is running A warning from Kennon tempers both theses: cost advantage only exists relative to competitors. In his hypothetical, improving your cost per task from 10 to 8 still leaves you more expensive than a competitor falling from 6. And current scarcity hides the difference: strong demand is expected to persist for two to three years, masking cost discipline the way it always has. His reminder is historical: every commodity in the history of humanity has had swings in supply and demand. His prescription is to build operating capability and cost advantage while capital is available, because it matters most when customers have choices and prices come under pressure. Long Lake's deployment playbook follows the same urgency through its own cycle — the J-curve of P&L transformation. Deployment starts as investment: platform build-out, workflow capture, close partnership, iteration, possibly training new models on the business's own data. Adoption then unlocks capacity — the same output in fewer hours — but capacity is not the finish line. Naidu's own standard: Don't just stop at capacity and time savings. How do you actually ensure that the work you're doing is driving real business impact. The measurable result is higher revenue per employee and better service quality for clients. A shared platform and playbook then compound across the portfolio, shortening both deployment time and time-to-P&L-impact with each new business. Key takeaways Value capture math: A $1M AI contract generating $5M of annual EBITDA impact creates ~$50M of enterprise value for the client at a 10x multiple — the structural reason vendors capture so little. Buy versus build: Buying an incumbent buys trust, distribution, veteran expertise, and live operational data that a from-scratch build must earn over years (the health-system architecture example). Margin structure: Services keep ~$20 of every $100 of incremental revenue versus ~$80 in software; AI breaks the revenue–labor link, an 80% marginal tax on growth falling away. Input versus output: GPU hours are input, intelligence is output: the same B200 stack can deliver ~50% more usable tokens through operator choices — rent per acceptable token is the metric that matters. Relative cost and urgency: Cost advantage is relative and cyclical: scarcity can hide weak cost positions for two to three years; build operating capability while capital is cheap, and treat capacity gains as the start of P&L impact, not the end. What you can now evaluate You can now put a price on each strategic option. When you evaluate buy versus build, weigh greenfield control against the upfront cost of trust, distribution, and tacit expertise — and remember the value-capture math that punishes outside-in vendors. When you evaluate M&A targets, look for mission-critical services with trusting customers, human judgment, and regulatory accountability, where owning the outcome — the urgent leak, the tax decision, the disrupted itinerary — is the product. When you evaluate cost levers, distinguish input from output: rent per GPU hour and headcount are inputs; price per successfully completed task and revenue per employee are the outputs that reveal whether operations are compounding or merely spending. Both talks, from services M&A to GPU operations, point at the same capability: the operator, not the asset, determines how much of the value you keep.