Last week, we saw the Old Guard of infrastructure get repriced in real time. Ericsson fell more than 12% after AI-inflated component costs ate its Q2. IBM fell 25%—a $70 billion wipeout, its worst day in 115 years—because clients are raiding hardware budgets to hoard AI servers. Looking down the row, Accenture is down 50% this year. The market has stopped paying for boxes and billable hours, and the value is collecting in the layers above. Ask yourself: do you own a layer that compounds, getting better as you use it, or a layer that commoditizes?
Read Benedict Evans’s latest essay—twice. His argument: nothing visible today gives foundation models durable pricing power, so they’re trending toward low-margin commodity infrastructure. His cautionary precedent? Telco. Mobile data grew into a trillion-dollar revenue industry with $200 billion in annual CapEx, and the stocks went nowhere; the value was captured up the stack. You lived that movie once. Now the model labs are cast in your old role, and the up-the-stack seat is finally open. Own your context layer this time. Don’t repeat the mistake.
And here’s proof that models officially became a commodity this month. In one 24-hour stretch, xAI shipped Grok 4.5 at $2 per million tokens, OpenAI shipped GPT-5.6 claiming it beats Opus-class models at a quarter of the cost, and Meta started charging for models it used to give away, with Zuckerberg attacking rivals’ “very extreme” margins on the record. Every launch pitched price, not smarts. When frontier AI labs compete on cost per token, buyers win, but only buyers who can switch. Never marry a model. Make the labs fight for your business every quarter.
Four years after CAMARA launched, network APIs are still a hot mess and McKinsey’s promised $100–300 billion is nowhere in sight. Analysts at DTW Ignite blamed missing “harmonization,” i.e. a developer building fraud detection needs “verify this SIM” to work the same way across 50 operators: same request, same answer, same meaning. Today they get 50 flavors. CAMARA standardized the doorway, but every operator has different furniture behind the door. Developers don’t buy connectivity; they buy outcomes they can build on. Until the industry agrees on what its data means, the plumbing stays worthless. Fix the meaning, then collect the money.
Akamai says agentic AI has a “latency crisis” and the cure it’s selling is compute distributed across its 4,000+ edge locations. Hmmm. When someone tells you the future requires thousands of edge sites, check who owns thousands of edge sites. AT&T sees it differently. Its CTO won’t chase the far edge “to save another millisecond or two,” and its play is AWS Interconnect – last mile: fiber and fixed wireless straight into the hyperscalers. (Catch the Shawn Hakl Telco in 20 episode where he talks about it.) Most operators agree: only 15% rank the far edge as the top spot for inference. AT&T has it right: you don’t need to own the compute. Instead, be the fastest path to it.
At DTW Ignite, Oracle’s Tony Gillick said that trust, confidence, and repeatability are must-haves before the industry can hand over business processes to AI. Careful with that framing: “building trust” usually means humans reviewing everything AI does, forever. If you cap your AI at human speed, human throughput, and human cost, you’ve automated nothing. The models are getting better by the day, and the Totogi Ontology makes wrong actions impossible to execute. When you build it right, you can trust your AI’s results.
Verizon and AT&T just put on a clinic. Verizon carried nearly 500 terabytes (TB) across World Cup stadiums—29.2TB during the final alone, uplink surging 70% at halftime—after laying 80,000 miles of fiber and upgrading venue capacity 3–5x. AT&T invested $430 million across 1,500+ projects and carried a petabyte at 99%+ reliability. Flawless. The economics? Verizon paid FIFA for a slot analysts peg at $65+ million, fans paid nothing extra, and the platforms monetized every uploaded goal celebration. In his latest essay (linked above), Benedict Evans described this: record traffic, value captured up the stack. You built the greatest network on earth for someone else’s highlight reel. Your next build should pay you. Dai dai, ikou, dale, allez, let's go!
