Jul 2, 2026, 10:00 AM ETInterview, Fireside Chat, Roundtable

Coinbase Cuts AI Spend by 50% | Kalshi's $40B Valuation & Impending IPO | The Year for SaaS Roll-Ups

Brian Armstrong (Coinbase) reported a 50% reduction in AI token spend over two months while usage (tokens generated) increased, attributing the efficiency to switching toward open-source models.
Jason Lemkin argues that "performative" AI data from struggling non-AI CEOs (e.g., Airbnb, Adobe) is unconvincing without tangible revenue lift; he demands to see specific growth acceleration tied to AI utility.
The consensus is that cost discipline is becoming the new normal for software companies, even those with strong balance sheets, as the "productivity lift" from agentic coding has failed to materialize proportionally to token spend increases.
Harry Stebbings notes that while the Coinbase memo is fact-based, it fails to address whether the 50% spend cut will reignite revenue growth in a volatile crypto market, highlighting the disconnect between cost-cutting and top-line expansion.
The discussion suggests that CFOs across the Fortune 500 are likely prioritizing AI ROI justification over aggressive experimentation, forcing companies to tie AI budgets to measurable Net Revenue Retention (NRR) or revenue acceleration.

A central tension exists between the $1T revenue ambition of frontier models (Anthropic/OpenAI) and the rise of cheap, open-source alternatives that could cannibalize market share.
Rory O'Driscoll warns that if companies optimize spend by 50%, frontier model providers like Anthropic could see their revenue growth rates decelerate significantly, potentially impacting their path to valuation.
The market is transitioning from an oligopolistic era (where competition was on features) to a commoditization phase driven by price erosion and open-source efficiency.
Harry Stebbings likens the US economy's dependence on AI to the "oil situation in the Persian Gulf," suggesting that regulatory capture may occur to protect the stock market and 401(k) valuations tied to AI infrastructure.
The speakers argue that while the government might be tempted to ban Chinese open-source models to protect US frontier model economics, the "open" nature of the code and lack of telemetry to China negates genuine national security risks, making a ban a "dumb" protectionist move.

Anthropic has formally alleged to the US Senate Banking Committee that Chinese open-source companies are illegally "distilling" IP by sending millions of prompts to record answers for training.
Jason Lemkin predicts a plausible future where US policy restricts Fortune 500 companies from using Chinese-origin models, not due to security, but to protect the high-cost infrastructure business models of US AI giants.
The speakers identify a potential "regulatory capture" scenario where frontier model providers leverage national security narratives to eliminate low-cost open-source competitors.
An alternative corporate strategy identified is making enterprises "uncomfortable" enough to voluntarily ban open-source models due to perceived (but unproven) security risks, effectively creating a self-imposed market restriction.

Microsoft shares have dropped to their lowest point since 2000, driven by investor skepticism regarding its lack of a standalone, state-of-the-art frontier model.
The market concerns center on Azure's growth deceleration (guiding 37% vs. expected 40%+) and the realization that much of Azure's AI revenue is merely inference reselling for OpenAI rather than proprietary product growth.
Rory O'Driscoll contrasts Microsoft with Google, which possesses its own standalone models, suggesting Microsoft's 30% equity stake in OpenAI is no longer sufficient to justify its valuation without its own "compelling AI product."
The deceleration of Azure is viewed as a "canary in the coal mine" for the broader AI economy, potentially triggering regulatory intervention to prop up the sector.

The volatility surrounding a potential SpaceX IPO is causing anxiety for AI startups like Anthropic and OpenAI, potentially delaying their exit strategies despite "greed trumping fear" in the current market.
Bending Spoons is going public at a $20B valuation (8-9x forward revenue), acting as a "roll-up" of legacy SaaS companies (AOL, Evernote, Notion), signaling that the IPO market remains active for companies with growth despite being "anti-AI" in nature.
Jason Lemkin proposes a "B2B Bending Spoons" thesis: acquiring underperforming, sticky-revenue B2B SaaS companies (e.g., Marketo, Asana, PagerDuty) and accelerating them via AI integration and improved customer success.
The speakers critique Private Equity firms for installing mediocre executives in acquired companies who fail to innovate, contrasting this with the aggressive, high-velocity cultural changes required to make legacy software viable in the AI era.

Harry Stebbings declined to back a startup finishing the year at $1.5M ARR with a projection of $5M next year, citing the "opportunity cost of cash" in an environment where top-tier deals are scaling to $50M+ ARR.
The community debate centered on whether Stebbings was too harsh; the conclusion was that while $1.5M-$5M growth can yield generational companies, it is insufficient for a traditional Series A in the current "AI-obsessed" capital climate.
Jason Lemkin criticized the lack of VC honesty with founders, noting that VCs often encourage fundraising processes they internally know will fail due to poor growth metrics, leaving founders with false hope.
Chamath Palihapitiya raised $135M for his new AI software factory platform, 8090; Lemkin expressed skepticism about his full-time dedication, suggesting he is not operating at the "insane rate" required of a traditional founder-CEO.

Anthropic launched "Claude Tag," an autonomous AI agent integrated into Slack channels (legal, product, etc.) that can act as a fully present team member, potentially disrupting traditional software workflows.
Jason Lemkin views Claude Tag as a potential "Trojan horse" that could render CRM platforms like Salesforce and HubSpot into "dumb databases" if the agent successfully captures the "context graph" of how work actually gets done.
The strategic implication is that if AI agents can autonomously handle context, data flow, and analytics across apps, the traditional "headless" software model may be upended in favor of a platform-agnostic agent layer.
Despite the existential potential, there is skepticism regarding whether Anthropic will prioritize such integrations if they represent less than 10% of their revenue stream, given the sheer scale of their operations ($100B+ revenue goals).

"Software companies in the age of AI are either accelerating or irrelevant."
"If you can be the largest tech company on the planet and still not make money, you might have oversized your ambitions a little."
"When elephants dance, the little people get trampled."
"Greed will still trump fear... right?" (Referring to the IPO market sentiment).
"The opportunity cost of cash is real." (Referring to the current VC funding environment).