Good morning. This was the week that the AI industry collided with physics—and blinked. If you only have two minutes, here’s what matters: the era of building your way to AI dominance through sheer scale is hitting a wall. The winners from here will be the ones who do more with less. Let me walk you through it.
Oracle and OpenAI quietly cancelled their massive Texas data center expansion this week. On its own, it’s a project cancellation. In context, it’s a turning point.
This was part of the “Stargate” initiative—the ambitious joint venture with SoftBank that was supposed to define the next generation of AI infrastructure. The kind of project that gets announced on stage with superlatives and renders. But behind the scenes, the financing negotiations dragged on, the grid capacity wasn’t there, and someone finally did the math on what 85 gigawatts of U.S. power demand by 2030 actually means in practice.
It means not every announced gigawatt will materialize. Hyperscaler capex has been exploding at 70% annually, reaching a projected $770 billion in 2026. Asia’s chipmakers alone are planning $136 billion in AI-related spending this year. But the Oracle cancellation reveals the gap between ambition and execution—between the keynote slide and the power grid. NVIDIA and AMD stocks dipped on the news, and that tells you everything about where the market thinks the constraint lives now.
If you’re building an AI roadmap, audit it for energy efficiency this quarter. As I wrote in 13 Predictions for 2030, power—not silicon—is becoming the binding constraint. This week proved it.
OpenAI shipped GPT-5.4, and it’s not just another model update—it’s an architectural statement. For the first time, reasoning, coding, and agentic capabilities are unified in a single model with native computer control, a one-million-token context window, and bidirectional audio for voice assistants. In practical terms, this means zero-code app development, autonomous workflows that chain across applications, and a 47% improvement in token efficiency. It scored 87.3% accuracy on complex spreadsheet tasks. That last number should worry anyone selling enterprise productivity software.
The benchmarks are even more striking. GPT-5.4 set a new state of the art on OSWorld—75% success rate on real-world operating system interactions, beating the human baseline of 72.4%. It also scored 88% on professional Capture the Flag cybersecurity challenges, which earned it a “high” risk rating. The business implication is clear: agentic systems that can navigate real desktops, fill in forms, switch between applications, and handle IT tasks are no longer theoretical. They’re here, and they could automate 30% more IT and security workflows than current tools.
Meanwhile, on the open-source side, Alibaba released the Qwen 3.5 series—spanning from 0.8 billion to 397 billion parameters. The 9-billion-parameter model is the standout: it outperforms rivals four times its size in visual reasoning, supports 262K context, and covers 201 languages. This is exactly the kind of small-model breakthrough we’ve been tracking in Chapter 12 of the report—fine-tuned specialists beating generalists at a fraction of the cost. The model runs on-device, which makes it relevant for global enterprise tools that need to work offline, in low-bandwidth environments, or under data sovereignty constraints.
The most revealing contrast this week was between OpenAI and Anthropic. OpenAI signed a Pentagon deal to deploy AI in classified systems—with surveillance bans attached, but a defense contract nonetheless. It won the bid where Anthropic lost. Revenue diversification into government is smart strategy, but it comes at a cost: a reported 295% spike in user attrition among consumer segments who chose OpenAI partly for its original non-profit mission. Brand erosion is hard to measure until it’s too late.
Anthropic went the other way entirely, publicly refusing military surveillance and weapons applications. It lost the contract but bolstered its ethical positioning in enterprise, where it reportedly holds an estimated 50% of corporate AI spend share. Principles as strategy. Both companies made the right move for themselves—but the divergence tells you that the AI industry is splitting into distinct identity lanes, and customers will increasingly choose based on values, not just benchmarks.
NVIDIA made a quieter but potentially more consequential move, building a Groq-powered inference chip. This is a shift from training dominance to cost-optimized inference hardware—exactly where the market is heading as enterprises move from experimentation to production deployment. And of course, the Oracle/OpenAI Texas cancellation we discussed above: Oracle loses scale but continues its 4.5 GW capacity build elsewhere, while OpenAI reevaluates its growth assumptions, opening a window for Meta to capture data center share.
Together AI released FlashAttention-4 this week, achieving 1,605 TFLOPs/s on NVIDIA’s Blackwell architecture. The significance isn’t the raw number—it’s the signal. When you can’t build power plants fast enough to feed the next generation of data centers, you make every watt count. Inference optimization is quietly becoming the most important discipline in AI engineering. AI data centers are projected to drive 20% of U.S. grid demand by 2030. FlashAttention-4 is the kind of software innovation that buys the industry time while the infrastructure catches up.
Block cut 40% of its workforce this week, citing AI-driven productivity gains. Read that again. Not “restructuring to invest in AI.” Not “reallocating headcount.” Forty percent of the workforce, gone, because AI tools made them redundant.
This is the clearest signal yet that agentic tools are reshaping the economics of software companies from the inside out. Cursor, the AI-native code editor, is now at $2 billion in annual recurring revenue—automating coding and support roles at a pace that makes traditional SaaS headcount models unsustainable. A viral essay on AI job replacement triggered an 800-point Dow drop this same week, capturing the anxiety in a single number.
But the anxiety obscures the opportunity. Firms adopting tools like GPT-5.4’s Codex are seeing 30% efficiency gains immediately. The real divide isn’t between companies that use AI and those that don’t. It’s between those that are AI-native and those still bolting AI onto legacy architectures. The latter aren’t falling behind. They’re becoming irrelevant.
Read the full SaaS disruption analysis →
Cursor Automations. If you write code—or manage anyone who does—this is worth thirty minutes of your time. Cursor’s new automation layer lets you build specialized agents for codebase analysis, Slack integrations, and timer-triggered tasks, all running on GPT-5.4. Think of it as autonomous coding workflows that actually work: you define the intent, and the agent handles the refactoring, the tests, the documentation. Teams are reporting a doubling in productivity. Install via JetBrains IDEs or standalone, start with a simple repo refactor, and judge for yourself. The speed difference is visceral.
“As AI’s grand infrastructure ambitions falter under funding and energy realities, the real intelligence advantage shifts to those who can do more with less—smaller models, smarter routing, fewer watts. The era of brute-force scale is ending. The era of orchestration has begun.” — Jose Antonio Martinez Aguilar
See you next Friday.