Coinbase Cuts AI Spend by 50% | Kalshi's $40B Valuation & Impending IPO | The Year for SaaS Roll-Ups
Summary
Coinbase’s 50% AI-spend cut is both a cost-control template and a warning about frontier-model growth. Rory O’Driscoll argues every Fortune 500 CFO will ask the CIO to replicate a factual result: token generation kept rising while spend returned roughly to November’s level after coding drove an explosion in November and December and companies took about five months to optimize. Jason Lemkin values the data but calls the surrounding CEO commentary “performative” unless AI produces revenue: “Show me the money.”
The next phase of enterprise AI requires measurable revenue lift or savings, not token-maxing. Jason’s strongest evidence is a top-0.5% portfolio company that hit plan but could not tie a proposed doubling of its already-massive token budget to ROI; the increase would have turned “no big deal” burn into a big deal. The board’s response marks the shift: “It’s time for the next mature phase of token spending.”
Cheap open-source models threaten Anthropic less as a product than as a trillion-dollar cost structure. Rory accepts that frontier models could retain most industry revenue even if open source generates most tokens, but alternatives at one-fifth the price become dangerous if Anthropic needs nearly $1 trillion of revenue to remain viable, as Harry frames Dario’s position. “If you can be the largest tech company on the planet and still not make money, you might have oversized your ambitions.”
Anthropic’s distillation complaint looks like groundwork for restricting Chinese models in the US. The alleged mechanism is millions of prohibited prompts whose outputs bootstrap competing open-source models; Rory distinguishes that contractual or copyright dispute from national security, saying offenders should pay a comparable “naughty tax” to Anthropic, analogous to what Anthropic had to pay book copyright holders, rather than face a ban. Jason nevertheless puts the probability of restrictions above 10%, while warning that merely making Fortune 500 buyers uncomfortable could achieve much of the same result.
Microsoft’s 16%-plus monthly decline reflects a software giant without a compelling owned AI growth engine. Rory says CoWork and Claude Code are attacking Microsoft’s historic knowledge-worker and developer franchises while Microsoft owns roughly 30% of OpenAI but lacks its own state-of-the-art frontier model. Jason focuses on Azure guidance falling from 40% to 37% growth: in a world supposedly moving toward “20 agents 24/7,” even that scale should be accelerating.
Kalshi’s proposed $40 billion valuation needs betting to expand far beyond political prediction markets. With roughly $2 billion of revenue and more than 70% reportedly tied to sports, Rory doubts it reaches $100 billion within 12 months unless it wins outsized sports share or financial products such as crypto perpetuals become enormous. Perps reduce investing to immediate up-or-down action — “Poor old Warren Buffett is like, ‘It’s time for me to die ’cause you people have lost the plot.’”
Bending Spoons could earn its premium by turning stranded software assets into a repeatable acquisition engine. Its expected $20 billion IPO implies roughly 8-9x forward revenue despite limited organic user growth, but Jason believes 1,000 identified targets could support five years of outlier growth. His B2B version would buy sticky, neglected products such as Marketo or PagerDuty, install operators who “give a crap,” and add AI-led revenue rather than merely repeating the 2021 private-equity cost-cutting playbook.
The venture bar has bifurcated: $1.5 million to $5 million ARR can build a great company, yet may not command a conventional Series A. Harry’s rejected founder appreciated hearing that reality, while Rory and Jason say most investors now compare such businesses with companies going from $1.5 million to $15 million or Higgsfield crossing $500 million in under 18 months. The honest prescription is not “go away and die”; it is lower fundraising expectations, potentially contact 150 investors, avoid a one-week process, and “cut your cloth accordingly.”
Deep dive
1. Coinbase turned token routing into a referendum on AI credibility
Harry’s starting fact: Coinbase cut AI spend 50% while increasing usage by routing more work toward open-source models and away from frontier providers. He asked whether this is the new normal or simply a frontier-minded founder moving unusually fast.
Jason accepted the chart’s value but rejected its framing as leadership: “I am getting burnout on struggling CEOs on Twitter sharing performative AI data when they’re not AI companies.” With Coinbase’s prior-quarter growth around minus 30%, he wanted evidence that AI improved the business, not proof that management optimized an input cost.
Rory’s pushback — worth keeping: Coinbase’s ordinary position between frontier startup and corporate America made the post useful. A “common or garden tech CEO” showed that any company spending $10 million or $50 million on Claude could keep generating more tokens while halving cost within two months.
His predicted distribution channel was the CFO suite: every Fortune 500 CFO likely forwarded the piece to a CIO with some version of, “Dude, look what this smart guy in the Valley is doing. Figure your shit out.” The post was cost management 101, not a promise to restore Coinbase’s growth.
2. Token-maxing is ending because boards cannot find the lift
Rory noted that Coinbase’s spend merely returned to roughly November’s level. Coding drove costs to explode in November and December, and it took about five months for companies to get their spend under control. For suppliers, a 50% customer cut is alarming even if spend returned roughly to its November level.
Jason’s deeper reading: product teams “radically ramped up” AI spend, saw qualitative benefits, but cannot show proportional productivity or revenue. “Jesus, I spent an extra 10 million in the first half of the year, and we grew the same as we did the prior two quarters. Like, where’s the lift, boys?”
One top-0.5% portfolio company made the problem concrete: every metric was green and the first-half plan was met, yet doubling an already-massive token budget would materially change burn. The board asked management to tie the request to ROI, and “this amazing team couldn’t.”
Jason’s standard is business-specific acceleration. Adobe claiming $500 million of agentic revenue while missing the quarter feels performative; Aaron Levie gets a pass because Box tied AI document processing directly to its model and returned to double-digit growth. “Software companies in the age of AI are either accelerating or irrelevant.”
3. Frontier-model economics are more fragile than frontier-model demand
Rory separated usage from value capture: open source could generate most tokens while state-of-the-art models still collect most revenue. Nothing in 100 Coinbase memos implies Anthropic and OpenAI cease being “amazing companies with great products that have differentiation.”
The risk is ambition. Rory recalled Anthropic’s run rate rising from roughly $1 billion to $9 billion last year and then toward $44 billion midyear; a hypothetical 50% haircut would not destroy the business, but it would affect growth. He explicitly did not predict revenue falling to $22 billion.
Harry framed Dario as saying Anthropic needs close to $1 trillion of revenue or risks bankruptcy. Rory’s response: even a half-trillion-dollar “consolation prize” would make it the largest digital company on Earth, so failure at that scale would mean the company “oversized” its cost and CapEx ambitions.
Rory’s bottom line: cheap alternatives at one-fifth the price are existential only if Anthropic “needs it all.” For a software customer, meanwhile, AI should produce either more revenue or savings; a company seeing neither will examine the bill “with a pretty jaundiced eye.”
4. Anthropic is converting distillation into a policy campaign
Rory began with the hypocrisy: foundation models trained on other people’s intellectual property, and Anthropic settled litigation with book copyright holders. “I do admire the element of hypocrisy, of being appalled when someone else does it to you,” he said, before turning to the substantive allegation.
Anthropic says Chinese open-source developers breached its terms by submitting literally millions of prompts, recording the answers, and using them as training data. That would let competitors bootstrap models from Anthropic’s output and then undercut the company with open-source substitutes.
Terms-of-service violations are initially contractual, Rory stressed: Anthropic can sue, though “good luck with the lawsuit, dude” in Beijing. Copyright or trade-secret claims could elevate the dispute into something the US government treats as more than a private commercial disagreement.
Writing to the Senate Banking Committee, whose minority head Rory identified as Elizabeth Warren, signaled serious political intent. Rory sees Anthropic “laying the pipe” for a trade: comply with US restrictions on overseas access, then prohibit US companies from using Chinese models proven in court to have distilled American technology.
5. A Chinese-model ban would protect AI producers by taxing every user
Jason believes Anthropic wants Chinese models barred from US corporate use and puts the probability above 10%; Harry asked whether it is closer to inevitable given support from both Sam and Dario. A partial win needs no formal ban: create enough security ambiguity that Fortune 500 companies prohibit the models themselves.
Rory separated punishment from security. Proven distillation should incur a comparable “naughty tax” paid to Anthropic, analogous to what Anthropic had to pay book copyright holders; it should not trigger exclusion. If the model is downloaded, its code is open for inspection, its weights are available, and nothing runs or sends telemetry back to China, he sees no national-security danger and “probably” would not ban it.
Jason’s contrary political economy: “AI is gonna be like the oil situation in the Persian Gulf today.” With 40% of the S&P 500 tied to the boom, politicians, 401(k)s, data centers, Nvidia sales, and employment all become aligned around protecting it: “Don’t touch my 401k.”
Rory called that protection “so fricking dumb”: it preserves expensive intelligence for foundation models while denying cheap intelligence to the rest of the economy. His analogy was banning Compaq and Dell clones to protect IBM and MS-DOS — keeping one incumbent profitable while preventing the PC industry from becoming enormous.
6. Open source is breaking the frontier-model oligopoly
Jason’s market history moved from what he described as two quasi-monopolies — Anthropic in coding and OpenAI in consumer — into an oligopoly where providers competed on features while prices, including roughly $200 maximum plans, remained broadly similar. The next stage is “massive price erosion” as competition shifts toward price.
He argued stable oligopoly pricing can coexist with feature competition and may be good for innovation in the short term, though perhaps not in the long term. Rory called that regulatory capture: oligopolies are excellent for participants precisely because they protect pricing, but “competition works.”
Rory’s Coinbase counterfactual made the consumer case. Without open-source alternatives, Armstrong’s post might have read: “We were paying $10 million six months ago for our AI intelligence. Now we’re paying 60. What the frick do I do?”
Any Chinese-specific restriction could still create an opening for US open-source suppliers such as Reflection and Poolside. Jason and Rory agreed policy is hard to predict; they disagreed on whether preserving frontier economics could ever justify suppressing the low-cost competitors forcing discipline.
7. Microsoft is being repriced for owning infrastructure, not the product
Harry described Microsoft’s worst month since 2000, down roughly 16%-16.5%. Rory’s explanation was structural: it has CapEx, cloud inference revenue, and a valuable OpenAI stake, but its core software business lacks a compelling end-customer AI product it owns.
CoWork and Claude Code are attacking the two historic Microsoft franchises Rory identified: software for individual knowledge workers and “developers, developers.” Unlike Apple, Microsoft is a software company and cannot remain outside the AI product war without losing the growth multiple attached to it.
Jason treated Azure’s guided deceleration from 40% to 37% as the canary. Markets now demand that companies “beat, raise, and grow”; if everyone will run 20 agents continuously, Azure should arguably accelerate despite the “law of gargantuan numbers.”
Rory added that much Azure growth is inference sold to providers such as OpenAI. Microsoft owns about 30% of OpenAI but lacks its own state-of-the-art model; since Satya Nadella said Microsoft would make Google “dance,” Google has massively outperformed and at least built a standalone model and product.
8. Kalshi’s $40 billion case rests on sports and financial gambling
Kalshi reportedly seeks a $40 billion valuation after raising at $22 billion in May and announcing roughly $2 billion of revenue. Rory’s simple explanation was that Americans like betting and the US’s longstanding restrictions on sports gambling have given way to a very large market.
More than 70% of the business is reportedly sports betting. Rory doubts Kalshi reaches $100 billion in 12 months unless sports keeps expanding without regulatory obstruction and Kalshi captures disproportionate share, or the non-sports financial side becomes far larger than expected.
Elections are entertaining but too small: relatively few people genuinely want to wager on the next president. The scalable human desires are sports and money; crypto perpetuals and simple up-or-down stock-price products turn financial conviction into immediate action.
Rory cited ICE’s roughly 20% interest in Polymarket as evidence that an established market operator sees prediction markets reaching major scale. His verdict on perps was admiration mixed with horror: “Poor old Warren Buffett is like, ‘It’s time for me to die ’cause you people have lost the plot.’”
9. Bending Spoons makes stranded products worth more together
SpaceX volatility has not closed the IPO market: Rory pointed to Bending Spoons’ planned July 1 listing as the “anti-AI IPO,” a roughly $20 billion buy-and-build company owning older products including AOL and Evernote. Jason said “greed will still trump fear,” though volatility could still make Anthropic delay.
Bending Spoons had roughly $1.5 billion of trailing revenue and $600 million in Q1, implying about 8-9x forward revenue. Rory found the premium strange because the assets reportedly add few organic users; management raises prices, cuts costs, and converts products formerly growing around 10% into a valuable roll-up.
Jason’s bull case rests on the prospectus claim that management has identified at least 1,000 material targets. If its acquisition and repackaging ability persists, he thinks it can sustain outlier growth for “five good years,” long enough to justify some premium even if the opening multiple is debatable.
Rory identified the acquisition arbitrage: many private consumer-software companies have no natural exit, making Bending Spoons the only buyer. The same logic should apply to subscale B2B companies that cannot meet a standalone IPO threshold of roughly $500 million revenue and 30% growth.
10. The B2B roll-up needs operators who rebuild, not merely optimize
Jason would start with nine-figure products possessing sticky customers and broken cultures. Marketo was his specimen: perhaps $300 million of remaining Adobe revenue, daily API failures, rate limits, a full-day outage, customer threats, and a promised 20% price increase without corresponding functionality.
His turnaround prescription is almost embarrassingly basic: install a motivated GM, stop threatening users, ship features, and preserve the base. Ten $200 million assets could become a $2 billion bundle growing 30%-40%, but only if Bending Spoons-style action begins quickly rather than after a new executive’s 90-day “learning tour.”
PagerDuty was the clearest second target: roughly 15,000 paying customers despite flat customer count and a market capitalization around $700-$800 million. Asana might qualify, though Jason warned that “agents don’t need Asana”; Rory also cited Semrush, bought by Adobe below two times revenue, as a missed GEO-bundling opportunity.
Rory’s pushback sharpened the thesis: Bending Spoons could optimize consumer assets, but pre-AI B2B software probably cannot survive by cutting costs and moving headcount overseas. It must generate new AI revenue and significantly re-engineer the product, making “Bending Spoons B2B” harder to manage but potentially more valuable.
11. Founder commitment, venture selection, and Claude’s software threat all converge on materiality
Chamath Palihapitiya raised $135 million for 8090, an AI software factory spanning new builds, refactoring, collaboration, and governance, while becoming CEO. Rory dropped the snark: Chamath is now “the man in the arena,” attacking the most exciting software market in decades, and deserves credit for trying.
Jason’s scar tissue produced a stricter test: wealthy investors running multiple projects often lose energy when “the S hits the fan.” He would invest only after Chamath dropped everything else and showed the physical and organizational scars of 100-hour weeks. “It’s too easy to start up today. Seeds for suckers, boys.”
The same selectivity explains Harry’s Series A rejection. Going from $1.5 million to $5 million ARR was top-quartile in old SaaS but is exceptional, not automatic, in AI-era venture; deals going from $1.5 million to $15 million get swept up, while Higgsfield reportedly crossed $500 million in under 18 months.
Yet slower growth is not failure. The founder thanked Harry for the candor; Jason advised speaking with perhaps 150 investors rather than forcing a one-week process. Rory’s correction was humane but economic: founders can still build generational companies, but must reduce the raise, approach a broader investor set, or converge on profitability.
Claude Tag applies that materiality test to software itself. An autonomous Claude embedded in Slack could absorb months of organizational context, move data across Salesforce and HubSpot, and make those applications “dumb databases” because “Claude is your head”; alternatively, it may prove little more than Zapier on steroids.
Jason sees no evidence yet that Claude Tag or Claude Design will sustain enough commitment to displace incumbents. Anthropic may soon be too large to pursue anything below $10 billion of revenue: Salesforce is roughly $42 billion and adds perhaps $8 billion, prompting Rory’s closing metaphor, “When elephants dance, the little people get trampled.”