☰
Blog · 22 Oct 2026 · Data as of 19 Oct 2026

Why Google and Meta are scored on 3 of 5

Next to every score on the board is a coverage label, like "5/5" or "3/5." It tells you how many of the five components (Arena, Artificial Analysis, revenue, valuation, Polymarket odds) actually went into that number. For Google DeepMind and Meta, the label reads "3/5," and that's not a data-collection gap the board plans to close. It's structural.

Why two of the components are missing

Both labs are units of much larger public companies: Google DeepMind of Alphabet, Meta Super­intelligence Labs of Meta Platforms. Alphabet's and Meta's quarterly filings disclose plenty, but not a revenue or valuation figure for the AI lab specifically.

The scoreboard's own data file says it directly. For Google: "This is a division of a public company. Alphabet does not disclose lab-level revenue or valuation, so those score components are excluded and the remaining weights are rescaled." For Meta: "This is a division of a public company. Lab-level revenue and valuation are not disclosed."

What Alphabet does disclose, at the Q2 2026 earnings release, is Google Cloud revenue of $24.8B (+82% year over year) and that the Gemini app reached 950 million monthly active users, as of 22 Jul 2026, filed with the SEC. That's a cloud-segment and a usage number, not a lab-level revenue figure, which is exactly the gap the scoreboard is pointing at. Meta's Q2 2026 results similarly give a company-wide 2026 capital-expenditure guidance of $130-145B (29 Jul 2026), again not a lab-specific figure.

The rescaling mechanic

When revenue and valuation are unavailable, their weights (20 and 10) are dropped, and the score is computed over the remaining 70 points of weight (Arena 35, Artificial Analysis 15, odds 20), instead of 100.

Google DeepMind, 19 Oct 2026: Arena component 100 (it is the Arena leader), AA component 91.2, odds component 79.8. Weighted sum: (35 x 100) + (15 x 91.2) + (20 x 79.8) = 3500 + 1368 + 1596 = 6464. Divided by 70: 92.3, matching the published score.

Meta, 19 Oct 2026: Arena component 85, AA component 96.8, odds component 4.3. Weighted sum: (35 x 85) + (15 x 96.8) + (20 x 4.3) = 2975 + 1452 + 86 = 4513. Divided by 70: 64.47, rounding to the published score of 64.5.

A middle case: Alibaba

Alibaba (Qwen) shows what partial coverage looks like when only one component, not two, is missing. Alibaba discloses an AI-related product revenue line (US$1.824B in the June quarter), which the board annualizes to $7.296B (x 4, per Alibaba's June Quarter 2026 results), so its revenue is scored, but it has no lab-level valuation (Alibaba is a public parent, so "valuation" doesn't apply the way it does to a venture-backed startup). Its coverage is "4/5," with 90 points of weight available (missing only valuation's 10).

Alibaba, 19 Oct 2026: Arena component 58, AA component 78.8, revenue component 65.9, odds component 0.7. Weighted sum: (35 x 58) + (15 x 78.8) + (20 x 65.9) + (20 x 0.7) = 2030 + 1182 + 1318 + 14 = 4544. Divided by 90: 50.49, rounding to the published score of 50.5.

The known bias

Rescaling means a lab never has to clear the revenue or valuation bar if it doesn't have one to clear. The methodology calls this out directly: "rescaling helps labs inside public giants (Google, Meta), which are scored only on models and odds. The coverage label is there so you can see it." It's a deliberate trade-off: the alternative, treating a missing component as a zero, would punish Google DeepMind and Meta for their parent companies' disclosure choices, which has nothing to do with model quality. The coverage label exists so readers can judge for themselves how much of each score to trust.

Sources

All blog posts →