By Damian Mathews & The Last Mile Team
On Monday, OpenAI confidentially filed for an IPO, with Reuters reporting a target valuation of up to $1 trillion and a debut that could come as early as September. Anthropic filed a week earlier. SpaceX prices its offering tomorrow and starts trading Friday at a valuation that would make it the largest public debut in history.
Add it up and close to $4 trillion in AI market value is heading for the public markets in a single summer.
Last week in No Hire, No Fire, we suggested reading the sudden softening of AI CEO rhetoric through the IPO lens. That aged quickly. Within days, both companies filed.
But the more useful story for CX leaders is sitting in the numbers underneath the filings.
OpenAI is reportedly losing $1.22 for every dollar of revenue it brings in. Internal projections show a $14 billion loss in 2026, with profitability not expected until 2029. The company is targeting a trillion-dollar valuation anyway.
That math has to resolve somehow. Inference costs have been falling fast for years: the price of equivalent performance drops on the order of 90% a year, and providers pass most of that through. The trillion-dollar bet is that the curve holds long enough to outrun the losses. If it doesn’t, prices rise. Probably some of both. And once these companies are public, the resolution stops being a patient engineering question and becomes a quarterly one.
Public companies behave differently than private ones. Earnings pressure shows up as pricing changes, enterprise sales pushes, product sunsets, deprecated model versions, and roadmaps shaped by whatever moves the stock. CX leaders have watched this movie with every major software vendor already in their stack.
The new part: your model provider joins the list.
If your contact center AI is built on a single provider, your roadmap is now downstream of their earnings calls. The agent you launch this fall could be running on a model whose pricing, deprecation schedule, and feature priorities are being set with an investor audience in mind by spring. Nobody at that earnings call will be thinking about your IVR.
The defense is architectural. Treat the model layer as swappable from day one. Abstract it behind your own orchestration so switching providers is a configuration change, a rebuild of nothing. Test your critical workflows against more than one model so you know your real switching cost before you need it. You don’t need an opinion on whether OpenAI is worth a trillion dollars. You need an architecture that performs the same whether it is or isn’t.
This is the posture we took internally. Every tool we documented in A1B: Customer Zero to AI-First was built assuming the model underneath would change, because it kept changing while we built.
And it’s the reason I can finally pay off the “more on that soon” we’ve been tacking onto the end of these issues the last few weeks.
This week we launched CX Foundry, a principal-led method Kerry (who you may remember from his hit weekly series Teaching Robots to Talk) has been building around the clock for months. It turns your real customer conversations into evidence, evidence into working agents, and simulated interactions into proof strong enough to make a launch decision. The tagline makes the connection to today’s argument better than I can: you don’t need a platform, you need a process.
CX Foundry defers the platform decision until the evidence justifies it, and everything it produces is portable. Run the proven agents in your own cloud, translate them to another platform, or have us run them for you. The full method is public, down to a markdown version you can paste into your own AI. One honest caveat… the site is days old and intentionally light on visuals. We shipped the content first because we’d rather have your feedback early than your applause late. Kick the tires and tell me what’s confusing.
The companies running the models in your stack are about to start reporting to Wall Street every 90 days.
What happens to your roadmap when your model provider misses a quarter?
— Damian
Here’s what went down this week.
Bleeding Edge
Early signals you should keep on your radar.
Anthropic released Claude Fable 5, its first Mythos-class model, days after warning that AI progress may be outpacing oversight. Fable 5 tops nearly every benchmark Anthropic tested, with safeguards that trigger in under 5% of sessions. If those guardrails hold up in production, cautious enterprise buyers may finally get frontier capability without frontier risk.
Apple rebuilt Siri on Google’s Gemini models and will let users route requests to Claude and other assistants through Extensions. The overhauled assistant handles multi-step tasks and arrives with iOS 27 later this year. Even Apple now appears to prefer renting frontier AI to building it, a signal every platform owner should weigh.
Leading Edge
Proven moves you can copy today.
Meta’s Business Agent went global inside WhatsApp and Instagram DMs, answering questions, booking appointments, and qualifying leads. Pricing folds into WhatsApp Business Premium tiers, with larger companies paying by token usage. Customer service is shifting into messaging channels at scale, and CX teams that wait may end up chasing it.
KPMG and Microsoft expanded their global partnership to deploy Agent 365 and Copilot across KPMG’s client engagements worldwide. KPMG will use Agent 365 to manage, monitor, and secure AI agents across client organizations. Agent governance looks set to become a service category of its own, and early movers could write the playbook.
Off the Ledge
Hype and headaches we’re steering clear of.
Meta shipped dormant facial-recognition code to more than 50 million phones inside its smart-glasses companion app, with no disclosure. The feature, internally called NameTag, built local faceprints and had lurked in the app since January. Meta pulled the code only after reporters found it, which says plenty about consent in the smart-glasses era.
Anthropic warned AI may soon build its own successors and urged the industry to preserve an option to pause development. Claude now writes more than 80% of the code merged into Anthropic’s own systems. A frontier lab asking for a brake pedal nobody has built yet should concentrate minds in every boardroom.
