Track Record

    The Board.

    As of Jun 2026·Re-reviewed monthly·methodology

    Think of this page as a chess board, not a crystal ball. Every call names the layer exposure a company is sitting on, the square under attack. The juggernaut still has a move. A master stroke (acquire L1, ship L3, lock the L5 workflow) can flip the ending. Most don't make it. That's the read.

    Each entry is scored on two independent axes: Structural (did the lens identify the right moat or exposure?) and Timing (did it arrive on the expected horizon?). Conflating those two is how frameworks lose credibility. Separating them, and naming the counter-move the subject could still play, is how this one earns trust.

    14
    Total calls
    9
    Structural confirmed
    4
    Structural playing out
    1
    Structural wrong
    5
    Faster than expected
    Filter
    Jump to

    The Chess Board of Intelligence

    Every company has the same 10 squares.The game is which ones they actually own.

    Think of the 10 layers as 10 squares on your side of a chess board. Some are queens (move in any direction, compound forever). Some are pawns (one direction, easy to lose). You play these squares against competitors and against the juggernauts, high-Elo players with money, compute, and frontier intelligence on the other side of the board.

    L-1 Resources
    Clock

    Not a piece, the clock on the wall. Decides how long any side can play. Megawatts, fab capacity, permits.

    Years to move
    L0 Infra5
    Rook

    Heavy piece, long lines of force. Hard to move once placed. NVIDIA, AWS, Azure, TSMC. Capex is the castling.

    Years to move
    L1 Data9
    Queen

    Moves in any direction, compounds forever. Bloomberg, Apollo, Tempus. Whoever holds the queen usually wins.

    Quarters to move
    L2 Models3
    Knight

    Jumps over other pieces. Powerful but swap-able, GPT, Claude, Gemini, Llama. Knight today is not the queen tomorrow.

    Weeks to move
    L3 Gates3
    Bishop

    Cuts diagonally across the board. Invisible until it pins the opponent, HIPAA, SOC2, audit, brand voice, indemnification.

    Years to move
    L4 Access5
    Rook

    Long lines down files and ranks. Chrome, Android, App Store, Salesforce. Hard to dislodge once it controls the channel.

    Quarters to move
    L5 Execution3
    Bishop

    Cuts across the workflow opponents can't reach. Harvey, Sierra, Cursor. The pin that turns a category vertical.

    Quarters to move
    L6 Orchestration3
    Knight

    Jumps over abstraction layers. MCP, LangChain, agent loops. Useful piece, rarely the one that wins the endgame.

    Weeks to move
    L7 Surface1
    Pawn

    One direction. Easy to lose, easy to ship. Jasper was a pawn that never promoted. Cursor is a pawn that did.

    Weeks to move
    L8 Memory9
    Queen

    The only piece that gets stronger every move. Sierra, Cursor, Clay. Memory that compounds is the long game.

    Quarters to move
    Rule I

    Juggernauts start with more material.

    Google opens with a rook on L4 (Chrome), a bishop on L5 (Workspace), a queen on L1 (search corpus). A startup opens with one pawn on L7 and has to promote it before the juggernaut castles.

    Rule II

    Some squares are hard to move into.

    L0 (fabs), L−1 (energy permits), L3 (regulator trust) are multi-year moves. L7 (surface) is a weekend. Time and money are the budget, pick the squares your clock can afford.

    Rule III

    The framework names the threat. The player chooses the move.

    A pinned pawn isn't dead. It has a counter-move: acquire L1, ship L3, lock L5, partner for L4. Every entry on this page carries the counter-move the subject could still play.

    Piece values borrow from chess convention (queen 9, rook 5, bishop & knight 3, pawn 1). They describe the layer's structural leverage, not the company's revenue. A pawn on a great file can still promote, Cursor did. A queen left undefended still loses, see BloombergGPT on L2.

    The Calls

    Every dated call, with the counter-move still on the table.

    1. Structural · ConfirmedTiming · On pace

      Sierra

      The call

      A vertical platform that owns L1 (per-customer policy + transcript data), L5 (workflow integration into existing CX stacks), and L8 (learning loop on resolution outcomes) is the structurally correct way to build durable agent-shaped value, and is what Jasper was not.

      What happened

      Sierra raised at $4.5B in late 2024 and $10B in 2025 on the back of named enterprise CX deployments and reported per-resolution economics. The L1/L5/L8 stack is exactly what acquirers and investors are paying a premium for.

      ♞ Counter-move · the master stroke

      Race Decagon to lock in the F500 CX logos with multi-year L1 (per-tenant transcript + policy) contracts; the L8 resolution loop only compounds if the L1 contract length outlasts Salesforce's bundling cycle.

    2. Structural · Playing outTiming · Too early to score

      Tesla vs. Waymo

      The call

      Autonomy is decided at L1 (real-world driving data) + L8 (fleet learning loop), not at L2 (model architecture). Tesla's data + fleet asymmetry compounds faster than Waymo's superior sensor stack + geofenced operational data, even though Waymo ships the safer product today. The framework's call: whoever owns the largest L1 + L8 loop wins the decade, regardless of who's ahead this quarter.

      What happened

      Through 2024–2026 Waymo expanded geographically with the better short-term safety record while Tesla's FSD v12/v13 made step-function progress on the back of fleet-scale data. The contest is unresolved and probably will remain so for years, exactly the multi-year horizon the L1/L8 call implied.

      ♞ Counter-move · the master stroke

      Waymo's counter: open the geofenced fleet data to third-party L5 partners (logistics, ride-share) to monetize L1 + L8 without waiting for full L7 consumer rollout. Tesla's counter: ship L3 (regulator-grade safety evidence) before FSD scales, the L1+L8 lead means nothing without an L3 license to operate.

    3. Structural · Playing outTiming · Faster than expected

      Devin (Cognition)

      The call

      A general-purpose L7 'AI software engineer' demo without L1 codebase ownership, L3 review/policy primitives, or L5 IDE workflow lock-in faces the same exposure as Jasper: the surface is reproducible inside Cursor, Copilot, and Claude Code the moment they decide to ship it.

      What happened

      Through 2024–2025, the agent-coding surface compressed: Cursor, Copilot Workspace, Claude Code, and Codex all shipped equivalents. Cognition pivoted toward Devin-as-teammate inside engineering orgs (an L5 move), validating the structural read.

      Timing noteSurface compression arrived within 9 months of the call.
      ♞ Counter-move · the master stroke

      Re-package Devin as an L5 teammate inside Slack/Linear/GitHub workflows with persistent L8 codebase memory, concede the standalone L7 surface, win the L5+L8 inside the org.

    4. Structural · ConfirmedTiming · Faster than expected

      Cursor

      The call

      Cursor owns L5 (the IDE workflow) and L8 (per-developer context / accepted-edit memory) in a way Copilot's bolt-on extension model structurally cannot match, the editor IS the workflow, and the workflow is where the agent loop accrues. The L5 + L8 stack consolidates the coding-agent category to whoever controls the IDE surface, not the model underneath.

      What happened

      Cursor's ARR scaled past $500M through 2024–2025 with developer adoption at the major AI labs themselves. Copilot Workspace, Claude Code, and Codex all repositioned around IDE-native or terminal-native workflows, explicitly conceding the L5 framing. The category consolidated exactly along the L5/L8 axis the framework predicted.

      Timing noteIDE consolidation arrived inside ~12 months of the call.
      ♞ Counter-move · the master stroke

      Keep widening the L8 moat (per-developer accepted-edit memory, team-level codebase context) and ship L3 (enterprise audit, code-review policies) so F500 procurement can't default-buy Copilot Workspace.

    5. Structural · Playing outTiming · Slower than expected

      Klarna

      The call

      Klarna's AI customer-service rollout is an L5 (workflow replacement) + L8 (per-resolution learning loop) move that could compress the L7 human-agent surface by ~700 FTEs without quality regression, validating in-house vertical-agent economics over horizontal Salesforce/Zendesk stacks.

      What happened

      Klarna's initial 2024 announcement claimed the cost win. In 2025 the company publicly walked it back, citing quality drift and CSAT issues, and began re-hiring human agents for higher-value tiers. The L5 compression was real; the L8 quality loop did not close fast enough to defend the all-AI tier. The framework's read on the economics was directionally right, the read on the quality-loop maturity was optimistic.

      Timing noteL8 quality loop matured slower than the L5 cost story. The two diverged in public.
      ♞ Counter-move · the master stroke

      Re-segment: ship a quality-graded L8 loop where the AI handles high-confidence resolutions and routes the rest to humans, with the routing model itself becoming the moat. Don't oversell the cost story before the L8 maturity catches up.

    6. Structural · Playing outTiming · On pace

      Glean (vs. Microsoft Copilot)

      The call

      Glean's L1 (enterprise connectors) + L3 (permissions/governance) + L5 (workflow integration) stack holds against Copilot inside non-Microsoft-monoculture enterprises; the bundle pressure is real but the L3 + connector breadth keeps Glean defensible where Microsoft does not own the substrate.

      What happened

      Glean continued enterprise expansion through 2024–2026 with raises at progressively higher valuations even as Copilot for Microsoft 365 went GA. Win-rate held in Google Workspace shops and mixed-stack enterprises; Microsoft-monoculture accounts remain contested as the call predicted.

      ♞ Counter-move · the master stroke

      Double down on non-Microsoft enterprises and ship deeper L3 (governance, audit, BYOK) primitives Copilot structurally can't match, turn the L3 gap into the procurement wedge.

    7. Structural · ConfirmedTiming · On pace

      Character.AI

      The call

      A consumer companion product whose entire value is L8 (per-user memory + relationship state) but whose L2 model is rented and whose L4 distribution is mobile-app store-dependent is structurally a 'memory orphan', the L8 asset is real but un-monetizable inside the cap-table constraints of a foundation-model-grade burn rate.

      What happened

      Google effectively acqui-hired the founding team in August 2024 in a structure that paid out the cap table without buying the company, the textbook outcome for an L8-rich asset trapped inside an L2-dependent cost structure. The L8 memory survived; the company did not.

      ♞ Counter-move · the master stroke

      Trade L2 independence for survival, sign an exclusive inference-cost deal with one frontier lab in exchange for compute credits, then monetize L8 memory via a premium tier before the cap table forces an acqui-hire.

    8. Structural · ConfirmedTiming · On pace

      Perplexity (vs. Google)

      The call

      Perplexity is a beautifully executed L7 answer-engine, but the framework's distribution law says L7 without L4 (default-channel placement) loses to whoever owns the query box. Google owns Chrome, Android, Safari-default, and the search slot, Perplexity has to fight for every install.

      What happened

      Through 2024–2026 Perplexity grew query volume meaningfully and raised at progressively higher valuations, but Google's AI Overviews + Gemini integration into Chrome/Android kept Perplexity's share of total AI-assisted search structurally capped. The framework's L4 distribution call held, owning the surface without owning the channel is a permanent ceiling, not a death sentence.

      ♞ Counter-move · the master stroke

      Acquire or build an L4 channel, browser, OS-level assistant, or default-search deal with a non-Google handset OEM, before Google's AI Overviews caps share permanently.

    9. Structural · ConfirmedTiming · On pace

      Stability AI

      The call

      An L2 foundation-model lab that open-sources its primary asset without an L1 data moat or L7 surface to capture demand has no place to extract margin, value flows to whoever owns distribution downstream.

      What happened

      Reported revenue / burn mismatch surfaced in 2024; CEO departure, board turmoil, repeated funding crises, and a 2024 rescue investment confirmed the L2-only-without-distribution thesis.

      ♞ Counter-move · the master stroke

      Pick a vertical (creative pros, gaming assets, advertising) and build an L1 (licensed data) + L7 (paid surface) stack on top of the open model, stop competing on L2, start competing on distribution.

    10. Structural · ConfirmedTiming · On pace

      Adobe Firefly

      The call

      Adobe's licensed-data L1 + Creative-Cloud L5 distribution gives it a structural moat against Midjourney and Stable Diffusion in the enterprise creative segment, even if the model quality (L2) is behind. Enterprises will pay for indemnification, not for the best raw model.

      What happened

      Through 2024–2026 Firefly became the default enterprise-safe image surface for Fortune 500 marketing teams, with the indemnification clause driving procurement decisions away from Midjourney despite Midjourney's quality lead. The L1 (licensed data) + L5 (Creative Cloud workflow) call held.

      ♞ Counter-move · the master stroke

      Defend by deepening L1 (more licensed catalogs, more language coverage) and tightening L5 (Creative Cloud-native workflows competitors can't reach), the moment Midjourney ships an enterprise indemnification tier, the L1 alone won't hold.

    11. Structural · ConfirmedTiming · Faster than expected

      Harvey

      The call

      A vertical legal AI that owns L1 (curated case-law + firm corpus) and L5 (workflow into associate / partner review loops) is structurally defensible against general-purpose models, because legal output requires citation-grade L1 and audited L5, neither of which a frontier model ships natively.

      What happened

      Harvey scaled to top-AmLaw deployments and a multi-billion valuation through 2024–2026. The L1 + L5 moat held, but the compression window narrowed faster than expected as frontier models (GPT-5, Claude 4.5/5) plus general legal agents from LexisNexis, vLex, and Thomson Reuters / Westlaw closed the citation-and-workflow gap. Harvey is still defensible; the lead it gets to keep is years, not decades.

      Timing noteFrontier-model + incumbent-stack pressure arrived ~12–18 months sooner than the original horizon assumed.
      ♞ Counter-move · the master stroke

      Lock in 3–5 AmLaw 100 firms with multi-year L1 (firm-corpus) + L5 (matter-management) integrations so the switching cost compounds before LexisNexis / vLex / Westlaw close the citation-and-workflow gap.

    12. Structural · ConfirmedTiming · On pace

      Chegg

      The call

      A homework-help surface (L7) whose only moat is an aggregated answer corpus loses its job-to-be-done the day an L2 foundation model can answer the same questions for free inside ChatGPT.

      What happened

      Stock dropped ~50% in a single session in May 2023 after CEO acknowledged ChatGPT impact on new-subscriber growth. Continued share-price decline through 2024–2026; subscriber base contracted year-on-year.

      Timing noteRepricing landed within the expected 12–24 month window.
      ♞ Counter-move · the master stroke

      License the answer corpus to OpenAI/Anthropic as an L1 supplier and pivot the surface to L5+L8 tutoring workflows (per-student memory, assignment-aware) before the L7 search interface collapses.

    13. Structural · WrongTiming · Faster than expected

      BloombergGPT

      The call

      Bloomberg's L1 (terminal proprietary corpus) + L2 (in-house trained 50B financial model) is the textbook vertical-fortress play: own the data AND own the model, and the L2 stays defensible against frontier-lab generalists in finance.

      What happened

      Bloomberg shifted its internal AI strategy toward using frontier models (GPT-4, Claude) over the proprietary BloombergGPT model through 2024. The L1 corpus remained the moat; the L2 ownership thesis did not, frontier-model rate-of-improvement made bespoke 50B-parameter vertical models economically and qualitatively obsolete inside ~18 months.

      Timing noteThe 'own L2' part of the call aged poorly. The 'L1 corpus is the moat' part survived. Logged honestly: the framework over-weighted L2 ownership against frontier-model improvement curves. This is the kind of call the lens has to get sharper on.
      ♞ Counter-move · the master stroke

      Concede the L2 race, double down on L1 (terminal corpus exclusivity, real-time market data licensing) and ship L5 (analyst workflows inside the Terminal), own the data + workflow, rent the model.

    14. Structural · ConfirmedTiming · Faster than expected

      Jasper

      The call

      An L7-only wrapper on GPT with no L1/L3/L5/L8 ownership compresses to zero margin the moment the underlying model ships a comparable surface.

      What happened

      Valuation reset from $1.5B (Oct 2022) to a fraction of that following ChatGPT, Copilot, and Gemini shipping native equivalents. Layoffs, CEO change, repositioning toward enterprise workflows in 2024–2025.

      Timing noteSurface compression hit inside ~6 months of the call.
      ♞ Counter-move · the master stroke

      Buy a vertical L1 corpus (industry-specific brand voice + compliance data) and bolt into a CRM L5 workflow before GPT-4 ships, convert from L7 wrapper to L1+L5 SaaS in under 12 months.

    Sources & citations

    Every call on the Board is grounded in publicly reported information. 42 citations across 14 companies, company posts (Primary), press coverage (News), stable reference pages (Reference), and live news searches (Search) for ongoing stories. Deeper sourcing lives inside each linked case study.

    Primary News Reference Live search

    Spotted a broken link or wrong source? Click the small flag next to any citation to report it, or see the Disclaimer page. Corrections applied promptly.

    On honesty, structural vs. timing

    A framework's job is to identify where value compresses and where it accrues. Its job is not to predict when - that depends on frontier-model release cadence, regulatory shocks, distribution deals, and cap-table accidents the lens does not see.

    So every call is scored twice. A structural call can be confirmed even when timing is faster (Harvey, Jasper, Devin) or slower than expected. Christensen, Porter, and JTBD all called direction correctly and timing wrong on multiple cases. Naming the variable the framework can't control is how it stays intellectually serious, and how it survives the cases it gets wrong.

    Calls stay on this page whether they age well or not. Anything that turns out wrong is marked wrong, not deleted.

    Disclaimer · editorial use only

    All company, product, and service names referenced on this page (and across Supply Chain of Intelligence™) are used descriptively, for editorial analysis, and remain the trademarks of their respective owners. Their inclusion does not imply endorsement, affiliation, or sponsorship, in either direction.

    Every reference is drawn from publicly reported information, press, earnings, primary documents, official blogs, and treated as a journalistic / blog source. Figures (valuations, ARR, headcount, funding) are reported as of the cited date and may change. We are not responsible for the accuracy of third-party reporting, nor for the real-world business outcomes, financings, or operating decisions of any company named.

    We are not ranking companies as "good" or "bad," "winners" or "losers." This is a lens, not a verdict. The framework identifies layer exposure - which squares a company is sitting on. The company itself decides what to do about it. That's the whole point of the counter-move.

    "The Chess Board of Intelligence" is a descriptive heading within this analysis. It is not claimed as a trademark.