You check CEO timelines the way you check horoscopes — fun to read, occasionally directional, never actionable. Here's a better oracle: a betting market where real people stake real dollars on whether a product ships.

The Polymarket contract for "Grok 5 released by June 30" opened at 68 cents last December. Musk promised Q1 2026 at the Baron Capital conference, and bettors — reasonably — took the over. Then February happened. xAI's official account quietly pushed the window to Q2. The contract slid to 40 cents. By late March, with all eleven co-founders gone and Musk admitting the company "wasn't built right the first time", it hit 20. Today: 12 cents. One-in-eight odds. The crowd has spoken, and the crowd brought receipts.

What makes prediction markets elegant is they digest information press releases refuse to. A CEO announces a timeline. A Polymarket contract adjusts that timeline against every leaked Slack message, every departing engineer, every quarter of radio silence. Tony Wu and Jimmy Ba walked out on February 10–11. The line moved. Manuel Kroiss and Ross Nordeen followed on March 27–28. It moved again. Eleven co-founders, zero remaining. No press conference needed — the contract did the math.

Meanwhile, bettors watched everyone else do the thing xAI apparently finds difficult: shipping. On April 2, Google released Gemma 4 — 31 billion parameters, Apache 2.0 license, outscoring models thirty times its size on math benchmarks. No gigawatt datacenter required. Weird how that works. On April 8, Anthropic launched Managed Agents — a full autonomous-worker platform built on existing Sonnet and Opus models, at eight cents per session hour. Less than the daily interest on xAI's hardware bill. A day later, MIT published CompreSSM, squeezing 4× training speedups out of model compression. Every one of these developments pushed Grok 5's contract lower, because every one proved brute-force scale stopped being the bottleneck when nobody was looking.

Colossus 2 — 555,000 GPUs, one gigawatt, roughly $18 billion — remains the largest AI training installation on Earth. It also remains the largest AI training installation on Earth with nothing shipped to show for it. GPU hardware depreciates at 30–50% per year at frontier pace; NVIDIA pushes new architectures annually. That $18 billion is a melting ice cube, and somebody unplugged the freezer.

If you're evaluating AI vendors for your product stack, here's the uncomfortable shortcut: the Polymarket contract for a company's next release tells you more about execution capability than any investor deck or CEO tweet. People with skin in the game price these contracts, update them in real time, and — critically — have zero incentive to be polite. Grok 5 at 12 cents is the most honest assessment of xAI's delivery timeline you'll find anywhere.

The AI industry spent 2024–2025 in a compute arms race. Prediction markets just scored 2026's first half: whoever ships, wins. Twelve cents says that's not xAI. And 555,000 GPUs in Memphis keep humming — still waiting for instructions from a team that no longer exists.