Alibaba’s LLM Qwen 3.8-Max became generally available this month with open weights and the highest open-weight score ever on Terminal-Bench (86.6%), beating both Claude Opus 4.8 and Fable 5 at several tests, at roughly a quarter of Opus’ pricing. This happens every few months: the frontier gets cloned and given away. Models are commodities now. That’s exactly why the Totogi Ontology sits outside the model, with your business logic, your constraints, your context kept in a layer ANY model can call. Qwen ships a better one Tuesday? Swap it in. The ontology stays yours.
AT&T’s internal Ask AT&T platform burns 45 billion tokens a DAY across 100,000 employees, generating a bill that would run in the hundreds of millions a year at frontier list prices. AT&T’s move, per The Information (paywalled): hold Anthropic and OpenAI spend flat and route the easy stuff—summaries, HR questions, call notes—to open-source models. Open-source models already handle 40% of queries, with AT&T predicting 70% in the coming years. A model router cut coding costs 56% with a 2% quality drop. That’s the playbook: your developers doing complex builds get Fable. Your call-note summaries get Nemotron. Paying frontier prices for commodity tasks is how you end up model poor: big AI bill and no budget left for the work that actually needs the horsepower.
To comply with the EU AI Act’s Article 50, Anthropic announced that all future models will embed invisible watermarks in the text they generate. What does the act accomplish? Not much. Anthropic admits the marks can be removed via translation or an edit; files’ metadata can be stripped through file conversion, screenshots, or re-saves. Bad actors making deepfakes will just grab an open-weight model (see above 👆) with no watermark and no one to fine. Brussels taxed every compliant AI company on earth and left the actual threat untouched. Classic EU: maximum burden, zero deterrence, and honest users’ work gets tagged while the fakes roam free. For my deep thoughts on this, see this week’s blog.
AI token growth is exponentially exponential. At Google I/O 2026, CEO Sundar Pichai said Google processed 9.7 trillion tokens a month in May 2024. Two years later: 3.2 QUADRILLION—a 330x jump. The wilder stat is who’s consuming them: agentic token usage surpassed human token usage on OpenRouter back in February, burning 15x the tokens per request. Goldman sees another 24x by 2030. Telcos, resist the urge to monetize this. I can already hear a strategy team somewhere pitching “token rating.” Stop. The labs sit above you in the stack and will own that layer, the same way over-the-top players owned APIs while telcos tried to charge for network access 15 years ago. You didn’t rate the API economy. You won’t rate tokens. Use them instead.
SpaceX closed its $60 billion all-stock acquisition of Cursor on August 14. It’s the largest startup purchase ever, paid entirely in shares of an IPO that was days old. Think about who’s making this bet: Elon owns rockets, satellites, spectrum, and the world’s biggest GPU fleet. He’s the ONE player with a credible shot at winning the infrastructure game outright—and the biggest check he’s ever written says the value isn’t there. It’s in the software layer. He’s armed with a public currency to keep buying too. If the king of infrastructure is racing up the stack, what does that tell you about a strategy built on staying at the bottom of it?
EchoStar chairman Charlie Ergen is spending $200 million for control of MobileX: an AI-powered, cloud-native MVNO that owns zero towers and right-sizes every subscriber’s plan with software. After a decade and over $30 billion spent chasing the fourth-carrier dream, he had Dish Wireless in bankruptcy, antennas stranded on towers, and Boost Mobile—by Ergen’s own earnings-call admission—“treading water” for four years. With MobileX, the combined business will have wholesale deals across all three national networks: infrastructure optionality without infrastructure CapEx. The man who bet biggest on owning a network just paid 1% of one spectrum sale to own a brand, an algorithm, and a customer relationship instead. It’s the industry’s future P&L in one transaction.
Recon Analytics finds a third of Americans can’t name their home internet technology—and of those who think they can, a third are wrong. After that opener, Fierce’s Linda Hardesty describes the myriad complications she encountered switching her family from T-Mobile to Verizon. Her conclusion: Because communication tech is complex, carriers need MORE human support agents. Wrong take. Because her port crossed a half-dozen systems that are integrated but don’t understand each other, she needed people to translate between them. An ontology like Totogi’s does that for you at machine speed, so the port becomes one autonomous flow: zero support calls, zero store visits. Customers won’t remember your network because you hired more people. But they will remember your friction. So, fix that.
