20VC: Dario and Anthropic Declare War on Open-Source | Coinbase Slash AI Spend by 50% | Kalshi's $40BN Valuation and Impending IPO | Bending Spoons: Smartest IPO of 2026 and the Year for SaaS Roll-Ups

Summary

Harry Stebbings and guests debate whether Coinbase’s reported 50% reduction in AI spend is a genuine signal that frontier-model usage is getting optimized away by open source, or simply a cost-management story that proves little on its own. That expands into a larger argument that open-source models, distillation, and even US policy responses could compress frontier pricing and reshape the AI business model. Microsoft is criticized for lacking a standalone frontier model and for relying too heavily on OpenAI and Azure inference revenue, while Kalshi is framed as a fast-growing consumer prediction-market business riding the normalization of betting. The second half turns to Bending Spoons as a template for software roll-ups: buying neglected products, raising prices, cutting costs, and potentially applying that playbook to B2B SaaS. The episode closes on the idea that autonomous enterprise AI agents could make traditional SaaS databases “dumb,” with Anthropic large enough that many individual vendor losses may no longer matter.

Notes

• Coinbase’s reported 50% AI spend cut is treated as a macro signal for frontier-model demand and open-source substitution.
• Speaker A dismisses non-AI CEOs posting benchmark takes as performative.
• Speaker B argues the Coinbase post matters because it is fact-based, with spend and token output shown over time.
• Cost cuts are not interesting unless they map to revenue, product velocity, or business-model impact.
• Speaker A says even strong software companies are now being forced to justify higher token budgets with explicit ROI.
• He says the market has moved from experimentation to a more mature phase where token spend must earn its keep.
• Open source is framed as a possible drag on frontier-model growth and economics.
• Speaker B says frontier revenue may still be dominated by state-of-the-art models even if token generation shifts toward open source.
• Anthropic’s legal posture is interpreted as trying to turn a distillation dispute into US policy.
• The alleged mechanism is that Chinese open-source players bootstrap via millions of prompts sent to Anthropic.
• Speaker A speculates that the US could ban Chinese models for US companies or impose a tariff-like penalty.
• The AI race is described as a national-security issue, with US policy likely to protect the boom.
• Speaker B pushes back, saying protecting frontier pricing would keep intelligence expensive for the rest of the economy.
• The market is portrayed as moving out of an AI oligopoly and toward price competition.
• Speaker B argues open-source competition is necessary to prevent AI prices from staying elevated.
• Microsoft is introduced as a company under pressure, with the stock down sharply.
• Speaker B says Microsoft’s AI strategy is weak because it lacks a standalone frontier model.
• Speaker A says Azure deceleration matters because the market wants acceleration even from huge platforms.
• Azure growth is also said to be partially driven by inference revenue for OpenAI and other providers.
• Microsoft is criticized for owning OpenAI equity without owning the model stack.
• Kalshi is framed as a $40B consumer prediction-market story, after a prior $22B round.
• Speaker B says Kalshi is mostly sports betting and is already strong on revenue.
• Speaker B says reaching $100B would require either far more sports-betting scale or much larger prediction products.
• Election prediction markets are described as too small to support a $100B outcome.
• Bending Spoons is presented as an anti-IPO: old software assets going public at a $20B valuation.
• Its economics are described as roughly $1.5B trailing revenue and about $600M in Q1, implying about 8.9x forward revenue.
• The value creation model is to raise prices, cut costs, and optimize tired consumer products.
• Speaker A says Bending Spoons could keep compounding because it has around 1,000 identified targets.
• He thinks a B2B version of the rollup model could work by fixing broken products and boosting NRR.
• The playbook is easier on old assets; B2B software would likely require actual AI-driven reengineering.
• Speaker A says declining public software companies should take even 15% premiums seriously.
• The current IPO bar is high, with standalone listings needing roughly $500M revenue and 30% growth.
• Speaker A says a 1.5M to 5M ARR path can still become a generational company, but it is no longer a strong Series A by default.
• Speaker B agrees the bar is tougher now, though some such deals can still work with extenuating factors.
• Claude in Slack is framed as a fully present, possibly autonomous member of a team channel.
• If that agent runs 24/7, it could become the control plane across apps and make SaaS systems just databases.
• Slack is valuable because it reveals the context and coordination patterns of work.
• Speaker A says Anthropic may become so large that many individual vendor losses will be immaterial.
• The episode closes on the image of a giant elephant trampling smaller players.