Podcast episode
Machines of Loving Grace? Kyle Csik on AI, Human Judgment, and the Future of Work
ai-in-adtech cloud-costs model-pricing
Signal & Noise hosts Rio Longacre and Brett House brought on Kyle Csik, founder and CEO of Adaly AI, to talk AI and the future of work. The episode's most useful argument has nothing to do with Dyson spheres or Turing tests.
Csik's operating point: the way most companies buy AI today, paying per token (each word or phrase processed by the model), gets more expensive as models improve. That's backwards from every software curve operators have lived through. His fix is federated querying, hitting your live data systems directly rather than warehousing everything first, so an LLM can actually answer a revenue question without hallucinating. His evidence for the crisis, including a $250 million token bill he pins on "I think it was Uber," is unverifiable, and he sells the solution he's describing.
The pricing structure concern is real, regardless of whether you buy Adaly. Check your own token invoices this quarter. That's the move, not a new architecture.
Full analysis
Your draft
Kyle Csik, founder and CEO of Adaly AI, spent an hour on Signal & Noise with hosts Rio Longacre and Brett House making one argument that matters to anyone running an ad-tech P&L: the way you buy AI today, per token, gets more expensive as the models get better, not less. Everything else in the episode, the Dyson spheres and the Turing test scores, is ambient. The token economics claim is the operating question.
The decision this forces on operators: keep renting intelligence from OpenAI and Anthropic by the token, or start pulling workloads onto open-source models you run yourself. Type 2, mostly reversible. You can pilot a local model without tearing out your frontier contracts. The forcing function is your next enterprise AI renewal and your own token bill, which nobody is watching closely enough.
The Market Analyst. Csik is selling federated data plumbing, so treat his frontier-model doom with suspicion. But the structural point survives the conflict of interest. Per-token pricing means your unit cost rises with model capability, which is backwards from every software curve operators have lived through. Plain version: normally software gets cheaper per unit as it scales, AI priced this way gets pricier as it gets smarter. Satya Nadella, in that TechCrunch piece from Sunday, told businesses that lean wholly on one proprietary lab they won't survive. When the CEO of the company selling you Copilot says don't bet everything on Copilot, the pricing anxiety is real, not a startup founder's talking point.
The Skeptic. The whole argument rests on whether open-source local models are actually a substitute for frontier models at ad-tech workloads, and that is far from settled. Csik cites Bayer, GSK, Tesla going open-source, and the summary flags every one of those as unverified. The $250 million single-quarter token bill he pins on Uber, he hedges himself with "I think it was Uber." So the whole cost panic rests on numbers nobody can check. For most ad-tech tasks, creative variation, bid explanation, audience summaries, a smaller open model may be plenty. But "may be plenty" is a bet, not a fact, and Csik profits from you believing it.
The Operator. Here is what breaks Tuesday morning. Csik's real point, buried under the philosophy, is that Copilot can't answer a basic revenue question because your data sits in silos. That is every ad-tech shop. Your spend is in one system, your delivery in DoubleVerify, your CRM in Salesforce, your analytics in Adobe. Bolt an LLM on top and it hallucinates because it can't see across them. The federated-query pitch, hitting live source systems instead of a warehouse, is the interesting operational idea. At 90 days the thing that bites is latency and permissions: querying five live systems per question is slow and a governance headache your security team hasn't signed off on.
The CFO. The token bill is a variable cost dressed as a subscription, and it scales with usage you don't control. That is the trap worth modeling. If your analysts start running agentic workflows that fire thousands of tokens per query, the bill compounds quietly. Open-source local deployment converts that to a fixed cost: buy the NVIDIA capacity once, run it flat. The catch is you now own the ops, the tuning, and the downtime. For batch work, creative generation overnight, reporting, local wins on cost. For anything real-time and customer-facing, the frontier API still earns its keep. Do the split.
The tensions. Csik and the Skeptic part ways on whether open-source is really ready. Csik treats it as obvious enterprise behavior; the evidence is all unverified and he's the vendor. The Market Analyst and the CFO agree the pricing structure is genuinely backwards, but disagree on urgency: Nadella's warning says move now, the CFO says model your actual token curve first because most operators aren't near the pain yet. And the Operator cuts against everyone selling architecture magic: bad data governance is the problem, and no federated query layer cleans that up.
What it hinges on. Two beliefs. First, whether per-token pricing actually rises fast enough to hurt your specific workload, which you can measure this quarter from your own invoices. Second, whether open models clear the quality bar for your batch tasks, which you can test in a weekend. Both are cheap to verify and neither requires buying Adaly or anyone else. The council leans toward: the cost concern is real, the vendor's cure is unproven, and the honest first move is instrumenting your own token spend before you architect anything.
No high-conviction prediction this week.
The episode is a founder's pitch wrapped in a philosophy chat. The pricing mechanism is worth internalizing, but there's no dated, checkable event here the source is plainly about. Predicting pharma model choices or Anthropic's valuation off unverified asides would be reaching past what this conversation actually settles.
Comments