Claude in Excel Is a Declaration of War on AI+X Startups

Anthropic released Claude in Excel, and the gap it exposes is not incremental. It is architectural. Microsoft uses AI to execute Excel functions. Claude uses Excel as a presentation layer. On both speed and quality, the architectural difference shows immediately, and that distinction (bolting AI onto existing software versus building AI-native from the ground up) is already separating companies that will survive 2026 from those that won’t.

Why “AI + X” Products Are Running Out of Road

Simply improving an existing service is no longer enough to hold market position. To compete seriously, a product needs to own at least one of three layers:

• Framework dominance: AI operating tools like Opencode and Clawdbot that define how agents work
• Infrastructure dominance: Foundational services like Vercel, Supabase, Cloudflare, and Stripe that everything else is built on
• Full-stack integration: Platforms like Base44, Replit, Lovable, Cursor, Comet, and Manus that combine the entire chain

If you don’t own any of these layers, your only remaining lever is price. A price war in AI products is not sustainable: costs scale with usage. More users mean higher fixed costs, not lower marginal ones.

Model Ownership Drives Step-Function Growth

Kuaishou, the company behind Kling, is generating $20 million in monthly recurring revenue. Google’s market share surged after releasing Nano Banana Pro and Gemini-3. These are not incremental gains. They are step-function jumps driven by model capability.

Then there is Grok. Largely overlooked in some markets, Grok’s penetration speed actually outpaces Google’s. xAI released world-class image and video generation models with virtually no restrictions, an aggressive play that traded safety guardrails for explosive adoption. Whether that trade-off holds long-term is genuinely unclear: regulatory pressure on unrestricted generation models is building, and the companies making that bet may face forced reversals.

The common thread is plain: the companies growing fastest own the model. Most startups do not, and building one is out of reach for nearly all of them.

Why Token Spend Is a Signal, Not a Cost Problem

“Why are we spending so much on tokens?” If you manage a team, you have heard this question. Until early last year, I did not have a great answer either.

Claude Code changed that. Watching it work, faster than most humans and producing output that rivals experienced engineers, made the value proposition concrete. You cannot understand what AI-native means unless you use AI intensively yourself. That lived experience is exactly where early-stage AI startups should be looking for direction. Rather than four separate bullet points, the insight collapses to one principle: find a narrow vertical with proven demand, engage deeply enough that you collect relationship data and workflow context your competitors cannot replicate, and design the product so AI is the architecture rather than a feature layer.

2026 Is Already Running Out of Time for Early-Stage Startups

One and a half months after this time last year, Manus launched. That company was later acquired by Meta for roughly $3 billion. Anthropic, Kuaishou, Vercel, Supabase, and Cloudflare are already preparing their next moves, and the window for a new entrant to establish position in a layer those companies have not yet claimed is narrowing each month.

The winners of the AI era will not be the companies that build better features. They will be the ones that define how AI itself operates, which means the most important question for any founder right now is whether their product could survive if Microsoft or Anthropic shipped the same capability tomorrow.