Welcome to Ignition, Catalyst Investors’ briefing on what we’re seeing at the intersection of AI, robotics, software, and growth equity, combining the best of human and AI expertise.
In this issue:
- The price war in Token Land
- Apple watches from the sidelines (for now)
01 — THE SIGNAL
The Price of Being Indispensable: The Story Behind the Coming Token Price War
There is a version of the AI price war story that is comforting to both OpenAI and Anthropic. Prices come down, demand expands, the market grows, and everyone wins. That version is incomplete.
The Wall Street Journal reported this week that OpenAI is considering drastic cuts to what it charges for tokens, in anticipation of similar moves from Anthropic. Sam Altman has said publicly that AI costs have become “a huge issue” for enterprise buyers, and that the company will find ways to deliver more value for less spend. Both companies are burning billions annually on compute. As they both prepare for IPOs, the market will want to see a path to margins that can justify their high valuations.
The timing follows Anthropic’s valuation overtaking OpenAI’s on the back of Claude Code, and now the competition has moved from model quality to price.
Why Prices Are Coming Down Regardless
On the supply side, the story is about Nvidia’s Blackwell GPU. SemiAnalysis compared Blackwell to its predecessor Hopper and found that measured by cost per million tokens, Hopper ran at $4.20. Blackwell runs at $0.12, a 35-fold cost reduction on the most important unit of AI economics. Blackwell systems are now being installed at scale, and by the second half of this year will be running the models for OpenAI, Anthropic, and Google. AI labs are about to have vastly more capacity to generate tokens cheaply. Price cuts are not just a competitive response; they are an output of a hardware cycle that has already happened.
On the demand side, three forces are converging. Foundation labs are moving up the stack into applications, competing directly with the enterprise software companies that are their biggest customers. Open-source models have crossed a “good enough” threshold for a meaningful share of production workloads. And enterprise buyers are actively routing around the most expensive closed models wherever they can.
The substitution is already happening. Lindy, the AI agent platform, switched 100% of its traffic to DeepSeek, citing millions in savings and better performance on core use cases. Harvey’s legal AI found that a fine-tuned open model outperformed Claude Opus on its benchmark at roughly one-eleventh the cost. Coinbase routes prompts dynamically to cheaper models, keeping total spend flat while token usage grows exponentially. The pattern is consistent enough across industries and company sizes that it is no longer just a series of anecdotes.
What makes this structurally different from a normal price war is the massive underlying demand elasticity. Enterprise buyers are reinvesting the savings from cheaper tokens in more AI. Business token usage grew over 1,000% from January 2025 to April 2026 while spend grew roughly 500%, meaning volume is running at roughly twice the rate of dollar growth. Goldman Sachs projects token consumption rising 24-fold by 2030. The cuts have already started. Anthropic reduced Claude Opus pricing 67% at the Opus 4.5 launch last November. A closely watched token spending index run by Silicon Data peaked in late May and has already started falling, which the firm’s CEO says may indicate prices are dropping across many models simultaneously.
The Margin Problem Is Real, But the ROI Problem May Be Bigger
The concern for investors is that two companies already losing billions annually are about to compete aggressively on price. The more interesting pressure point is whether enterprise demand growth holds up.
The data are mixed. MIT’s 2025 GenAI Divide study found 95% of enterprise generative AI pilots delivered no measurable profit within six months. Uber consumed its entire 2026 AI coding budget in four months, now caps per-tool spending at $1,500 per engineer per month, and its COO said publicly that drawing a direct line from AI spend to shipped features has become difficult to justify. The FinOps Foundation found companies running agentic workloads had already burned through three times their annual token budgets by April. On the other hand, only 6% of executives say they would cut AI investment even if 2026 return on investment disappoints. The enterprise-level ROI case is still in process, but no one wants to be left behind.
The true bear scenario, then, is not that model margins compress. It is that ROI skepticism spreads fast enough to slow the demand growth that would make margin compression survivable. Lower prices only work as a strategy if volume responds. If enterprise buyers conclude that twice as many tokens at half the price still do not pencil out, the volume math will break.
What Ripples Through the Stack
Anthropic has already lowered its gross margin projection to 40% because of inference costs running on Google and Amazon infrastructure. That margin gets harder to defend as prices fall further. For both companies, the margins public markets expect require either dramatically cheaper compute or pricing power their customers are actively dismantling. Blackwell solves part of the compute cost problem. It does not solve the pricing power problem.
The pressure accelerates the shift to custom silicon. Google TPUs, AWS Trainium, and Microsoft Maia each cut inference costs 30% to 50% versus Nvidia. TrendForce projects custom chips running 40% of AI servers by 2030. The token price war may be the event that makes that transition urgent rather than merely interesting.
The neocloud exposure is the most structurally precarious piece. Firms like CoreWeave carry fixed debt service against revenue streams tied to AI infrastructure demand. CoreWeave holds $29 billion in total liabilities, with $9.7 billion due within twelve months, and pays 9.75% interest on its senior notes. Oracle carries over $130 billion in debt plus $248 billion in lease commitments. If price cuts slow the AI capex cycle, the cash flow assumptions behind hundreds of billions in data center financing get tested for the first time.
What Matters for Investors
The price war, the margin pressure, and the ROI uncertainty are real. So is token usage volume, which, at least so far, keeps increasing more than enough to offset the price declines.
Cutting prices will make OpenAI and Anthropic less profitable in the near term. The question is whether the expansion of addressable workloads that cheaper tokens unlock compounds faster than the margin compression they create. That is a bet on elasticity. The historical pattern from technology markets that have gone through this transition (mobile data, cloud computing, bandwidth, etc.) suggests elasticity usually wins. So now both companies head into a pair of mega IPOs in a market that will want certainty neither company can provide. Tokens are going to be dramatically cheaper. The open question is whether that is a feature rather than a crisis.
Sources: The Wall Street Journal / Keach Hagey and Berber Jin (June 10, 2026). Tomasz Tunguz / Theory Ventures (June 7, 2026). Forbes / Peter Cohan (June 11, 2026). Business Insider / Alistair Barr (June 12, 2026).
02 — OPERATOR’S EDGE
What Apple Knows About AI That Silicon Valley Won’t Admit
An irreverent take by Alberto Romero — The Algorithmic Bridge (May 30, 2026)
We flagged the tortoise-and-hare framing last issue in the context of Google. This week at its WWDC (worldwide developer conference), Apple continued to say that it will sit this race out for now. Romero’s piece is an intellectually honest account of what Apple’s AI strategy implies, and it leads somewhere Silicon Valley is not ready to go.
The core argument:
- Apple is spending 2% of what its peers are spending on AI capex ($14 billion projected for 2026 versus roughly $670 billion combined across Amazon, Google, Meta, and Microsoft) and has made no attempt to close that gap
- Tim Cook’s successor is John Ternus, a 25-year hardware veteran with no AI background. Cook could have named an AI-first CEO. He did not. Romero reads this as institutional conviction, not oversight
- Apple is opening Siri to third-party models including ChatGPT and Claude, building a routing layer rather than a proprietary model. You do this if you believe AI models are a commodity, not a moat
- The hyperscalers are not true believers either — Each is making Pascal’s Wager: that the downside of missing out on AI and being wrong is worse than the cost of the bet, so they tithe to the church of AI. According to Romero, none of them really believe in what they are spending:
– Meta pours billions into whatever technology is hot at a given moment, from 3D printing to the metaverse to AI. Its primary consumer-facing AI product is a chatbot inside WhatsApp
– Google announces a 24/7 personal agent and an AGI seed in the same breath, then ships Gemini into Gmail where it suggests slightly worse replies and into Search where it hallucinates
– Microsoft invested in OpenAI before anyone else, then forces its enterprise clients to use Copilot instead
– Elon Musk builds a giant AI datacenter and rents it at a profit to a competitor
– Sam Altman believes in everything at different times — whatever the public wants to hear, he obliges
The market takeaway from WWDC this week is instructive, however. Apple’s stock fell 8%. Romero is not offering a consensus view on what Apple’s AI posture means for the long-term health of the business.
The Catalyst Take
The Romero piece hits at the same underlying question as the price war story in The Signal above: if AI models are becoming a commodity, who wins? Apple’s bet is that the value lives in the device, the identity layer, and the user relationship, not the model. The new CEO is a hardware guy, potentially preparing to launch the new hardware ecosystem for the AI age.
If that is right, the companies spending $670 billion on infrastructure are building roads that Apple gets to drive on for (almost) free. That is not a guaranteed outcome. WWDC showed Apple is still years behind on agentic capability, and the limited rollout geography (US only, no Europe, no China) is a real constraint. But Romero’s framing is a useful corrective to the assumption that more capex equals more defensibility. The asset-light position looks weak until the infrastructure cycle turns. Then it looks like it was the only rational strategy all along.
Forward to a founder, operator, or investor who is navigating AI adoption.
ir@catalyst.com
261 5th Ave
Suite 1102
New York NY 10016
