Cheap Decisions, Fatter Margins
Alex Pedraza@apedraza
The idea
I think the next wave of AI savings comes from swapping expensive LLM calls for cheap, narrow decision models on simple yes/no and labeling tasks, the kind of classifying and routing work that happens millions of times a day. To play it, I own software companies whose AI products get cheaper to run, transaction classifiers in accounting and finance, the platforms hosting these smaller models, fintech and security companies making huge volumes of these calls, and big cloud providers as steadier ballast, while staying away from GPU makers since cheaper inference works against them.
What will be bought
- These software companies run large AI bills to power their products, so cheaper inference costs flow directly to operating margins. · ServiceNowRoutes enormous volumes of IT and support tickets through AI agents, and routing is a pick-one decision where cheaper calls make agents more profitable., SalesforceSalesforce AI agents make many calls per customer conversation, and plenty of them are simple routing and yes/no checks where cheaper decision models cut the bill., HubSpotSells AI features to small and mid-size businesses where inference cost per call matters a lot to margins, and classification of leads and emails benefits directly from cheaper models., FreshworksSells help desk software where every incoming ticket gets triaged and routed in a natural classification task, and lower AI costs help smaller companies more than giants., DuolingoRuns a large AI bill for grading student answers, which is classification work, so cheaper decision models for grading turn more of each subscription into profit., TwilioCarries huge volumes of messages and calls that need sorting by intent and urgency in classification tasks, and cheap decision calls make that sorting something it can sell on every message.24%
- Accounting and finance software labels millions of transactions automatically, which is exactly the narrow classification task that decision models handle cheaply and well. · IntuitQuickBooks sorts millions of small-business transactions into categories, the same labeling work that moves to decision models as costs fall, and Intuit's growing AI bill makes cheaper inference valuable., BILL HoldingsEvery invoice and card swipe gets coded to an account in a pick-one question repeated millions of times, so cheaper decision models keep more of each dollar BILL processes., WorkivaTags financial reports with standard regulatory labels, which is a labeling problem at its core, and cheaper decision models lower the cost to serve each customer., Vertex IncAssigns the right tax category to every product and invoice line, a narrow labeling job that decision models handle better than large language models., BlackLineMatches and reconciles transactions during month-end close at huge volume, a classification problem where cheaper decision calls let it run more matching for less.20%
- These companies run and monitor inference at scale, so they benefit when more AI calls flow through their infrastructure at lower cost. · CloudflareRuns AI models on servers close to users and hosts the kind of small, fast models that grow as cheaper inference spreads across the internet., DatadogCompanies swapping larger models for decision models need to monitor cost and accuracy per call, and Datadog sells that observability as call counts climb., Arm HoldingsDesigns the chips behind many servers at Amazon, Google and Microsoft, and small decision models can run on low-power Arm-based processors instead of expensive GPUs.15%
- The major cloud platforms host every kind of model and win from higher usage volumes even as prices per call decline. · MicrosoftAzure hosts a wide range of models and gets paid on usage regardless of which model companies choose, so it benefits from higher call volumes even as per-call prices fall., AmazonAWS sells model access through Bedrock and runs its own chips, so cheaper inference brings more calls onto the cloud and higher utilization of its infrastructure., AlphabetRuns AI on its own chips and sells small, cheap models, so it benefits as companies chase lower inference costs and more calls flow through Google Cloud.15%
- Lending and payments companies approve, score and flag millions of transactions daily, so cheaper decision calls improve their unit economics and margins. · BlockScores Square and Cash App payments for fraud risk in yes/no decisions on every transaction, so cheaper inference lowers the cost of each payment it handles., UpstartThe entire business is AI deciding whether to approve a loan, which is exactly the narrow kind of yes/no call that decision models handle, and cheaper decisions improve cost per loan., PayPalRuns a risk check on every checkout in a high-volume yes/no decision, and moving those to cheap decision models lowers the cost per transaction PayPal processes.10%
- Security tools make billions of yes/no decisions like threat detection and traffic blocking, the exact work that decision models do cheaply at scale. · CrowdStrikeSorts enormous numbers of security events into safe or threat categories in yes/no calls, so even small drops in cost per call add up across billions of checks., ZscalerDecides whether to allow or block web traffic for enterprise customers in billions of yes/no calls, and cheaper decision models let it run more checking for the same cost.8%
- Cash8%