Rating Actions

    A dated record of score, evidence, listing, release, and catalog eligibility changes in the Korea AI database.

    2026-10-27

    rating

    HBM suppliers: comparable Adoption evidence
    Adoption: SK hynix 75 / Samsung 75 / Micron 96.3
    95 / 92.4 / 88.7

    Adoption now follows matched Q2 2026 HBM revenue shares: SK hynix 95, Samsung 92.4, Micron 88.7. Value 84 and Durability 82 remain unchanged.

    Counterpoint estimates revenue shares of 50%, 33% and 18% for the same quarter and market. A scoped, fixed logarithmic curve replaces mixed production fallbacks and Micron’s September 2025 run rate. That run rate remains historical evidence, excluded from this rating. Scores are editorial estimates with overlapping ranges, not measured quality differences or dollar revenue.

    2026-10-27

    rating

    Samsung and SK hynix HBM
    Samsung 74.5 / SK hynix 82.8
    Both 80.2

    Both HBM families now receive Value 84, Adoption 75 and Durability 82: Strength 80.2.

    Commercial HBM4 sales correct Samsung’s deployment evidence. Comparable quality, yield or exclusive-access evidence did not justify SK hynix’s additional Value/Durability points. Both remain estimates in the same rubric band, not a claim of equal quality or market share. Selected earnings releases do not establish isolated annual HBM revenue.

    2026-10-27

    rating

    1M monthly active users and $25M annual product revenue / ARR both receive Adoption 63. Six counted-use ratings are recalibrated.

    $25 per MAU per year is an editorial scoring equivalent, not observed revenue or ARPU. Registrations and downloads receive no active-user credit. Daily/weekly counts retain their cadence as monthly lower bounds without a multiplier. Source dates, confidence, enterprise revenue, Value and Durability are preserved.

    2026-10-27

    catalog

    18 sourced products cover 17 previously profile-only companies; four ineligible listings are archived. No profile-only companies remain.

    Each retained company owns a canonical product row with explicit estimated ratings and buyer comparisons. Research-only or inaccessible offerings are not invented to retain a profile. Original records and removal reasons are archived; retained financial figures are unchanged.

    2026-10-25

    evidence

    Twelve additional revenue observations; nineteen Adoption scores change. Enterprise listings now lead with revenue and dates.

    Twenty-four products received focused financial-source checks; seventeen enterprise records now have usable product revenue, ARR or run rates. Lunit separates INSIGHT and SCOPE. VUNO's DeepCARS is a suite-component floor; Backend.AI is a rounded estimate from disclosed product revenue shares. Seven unqualified customer-count scores fall to qualitative traction. Paid subscriptions remain a fallback. Company totals, funding, forecasts and undated ARR claims are excluded. The remaining catalog is not claimed to have a completed revenue search. Value, Durability and consumer ratings are unchanged.

    2026-10-25

    rating

    Equal geometric mean · strength-v9
    Weighted geometric mean
    ∛(Product Value × Adoption × Durability)

    Product Value, Adoption and Durability now count equally. All displayed scores remain on 0–100.

    Divide each component by 100, multiply the three values, take the cube root, then multiply by 100. Fixed endpoints preserve the meaning of each score when the catalog changes. All 1,262 product and 601 company snapshots are recalculated; component judgments, evidence and review dates are unchanged. The same formula combines uncertainty bounds and company portfolio components.

    2026-10-25

    rating

    Durability after feature cloning
    Lunit INSIGHT 72 · Backend.AI 69 · MarqVision 67 · Wrtn 32
    60 · 45 · 45 · 32

    Reviewed all 1,262 records: 15 focused source reviews and 635 catalog-based corrections. 646 Durability ratings change.

    A copied interface does not reproduce clinical validation, reliable operations or enforcement access. But ordinary integrations, customer-owned data and specialization do not prove strong lock-in either. Global products receive the same scrutiny. Adoption and historical cost reviews are unchanged; catalog-only corrections retain source dates. SCOPE’s Value comparison is corrected from molecule generation to pathology, raising 55 to 68. All totals use the existing three-part geometric formula.

    2026-10-25

    rating

    Three criteria, multiplied · strength-v8
    Four-part weighted sum
    Value^0.40 × Adoption^0.35 × Durability^0.25

    Product Strength now uses the weighted geometric mean of Product Value 40%, Adoption 35% and Durability 25%.

    A weak component limits the total more than in a sum. All product and company scores and ranges are recalculated from unchanged component judgments and evidence. Economics leaves the score; pricing, inference costs and disclosed profitability remain supporting information. Costs cannot be derived from the three ratings, and unknown margins stay unknown.

    2026-10-25

    rating

    Durability weighting and component order · strength-v7
    Durability 15% · Economics 20%
    Durability 20% · Economics 15% · 100% total

    Product Value 35%, Adoption 30%, Durability 20%, Economics 15% now determine Product Strength and the component display order.

    The maintainer transfers five percentage points from Economics to Durability. All product totals, uncertainty ranges and company portfolio scores are recalculated. Component ratings, source evidence, confidence and review dates remain unchanged. Tables, score panels and comparisons follow decreasing weight.

    2026-10-23

    catalog

    Added 10 companies and 14 products; expanded 23 product and 20 company buyer briefs.

    The same four-part rubric applies to every region. Availability, qualitative traction and named deployment remain distinct; partner logos, funding and nationality earn no bonus. New judgments are labeled estimates with named alternatives and broad ranges. Solar and RNGD keep their existing dated Adoption evidence. New in the catalog does not mean newly founded.

    2026-10-23

    catalog

    Gemini Notebook, Workspace, business agents, Code Assist and the enterprise agent platform join the Gemini entry.

    Eight components retain separate product links, billing context and recorded history. The representative app assessment stays at 89.4, with unchanged evidence; audiences are not summed. Google Cloud remains a separate company and links this shared family without counting its consumer app rating again.

    2026-10-23

    rating

    More weight on Product Value · strength-v5
    Adoption / Value / Economics / Durability: 30 / 25 / 25 / 20
    30 / 35 / 20 / 15 · 100% total

    Product Value rises to 35%; Economics falls to 20% and Durability to 15%. Adoption remains 30%.

    The maintainer places more emphasis on usefulness for the intended job. All product totals and uncertainty ranges, and company portfolio scores, use the revised weights. The four product ratings, supporting evidence, confidence and review dates are unchanged. This changes the weighting policy, not the underlying assessments.

    2026-10-23

    rating

    Economics on the full rating scale · strength-v4
    Economics maximum 78 · median 60
    Economics maximum 96.4 · median 70

    Strong Economics ratings can reach the 90s, on the same scale as the other criteria.

    Fixed anchors recalibrate 1,030 product judgments and their uncertainty bounds. The conversion does not depend on the catalog's top product or rankings. Adoption, Value, Durability, source facts, dates and confidence are unchanged. Unknown and adverse economics retain their earlier treatment. Economics is a delivery-efficiency rating, not a margin percentage or measured saving; weights stay 30/25/25/20.

    2026-10-23

    catalog

    Moloco advertising products

    Moloco Ads, Commerce Media and Streaming Monetization are excluded from the catalog by maintainer direction.

    AI is secondary to the advertising offering for this catalog's intended scope. The company profile and old product aliases are also removed because no qualifying products remain. Records are archived. This is an editorial listing decision, not a claim that the company closed or does not use machine learning.

    2026-10-23

    rating

    Adoption with diminishing returns · strength-v3
    Claude Adoption 85 · Strength 80.3; ChatGPT Adoption 95 · Strength 83.3
    Claude Adoption 94.9 · Strength 83.2; ChatGPT Adoption 99.4 · Strength 84.6

    Established use earns most adoption credit; reaching incumbent audience size adds progressively less.

    The same curve recalibrates every product: 34 Adoption point estimates change, while Value, Economics and Durability remain unchanged. Recorded counts interpolate smoothly; organizations and paid seats use separate thresholds. A million active users earns 94.9, a billion 99.5. Qualitative deployment categories use the curve without inventing user totals. This changes scoring policy, not source facts, dates, confidence or retention evidence.

    2026-10-23

    rating

    Claude, ChatGPT and specialist apps
    Claude app 72.8 · ChatGPT 85.2 · Harvey 76.1 · Glean 69.1
    Claude app 80.3 · ChatGPT 83.3 · Harvey 69.5 · Glean 64.2

    Claude matches ChatGPT on usefulness; legal and knowledge apps must justify their advantage against configured general assistants.

    The two general apps share Value, Economics and Durability estimates; their headline gap is Adoption. Customer files, retrieval and ordinary connectors receive less durability credit across 53 legal, contract and enterprise-search products. Harvey Tenet and Glean’s vendor comparisons remain bounded evidence, not proof of lasting superiority or product margin. The weights are unchanged.

    2026-10-23

    rating

    All 1,263 products now have four explicit editorial ratings and one weighted score.

    Adoption 30%, Product Value 25%, Economics 25% and Durability 20% replace the prior public metrics and unassessed placeholders. Estimated judgments remain distinct from observations; dated ratings stay visible after 90 days with a review notice. Company scores use the same canonical-product portfolio and never allocate company value. This supersedes the earlier withholding and expiry policy below.

    2026-10-22

    rating

    Full-catalog evidence audit
    Default scores, company/country priors and profile fallback
    Dated assessments or Not assessed

    Unsupported ratings are withheld; every published metric needs its own dated evidence.

    The audit covers every product and company. Impact is a deployment band, Global records reach, and commercial ratings use customer and cost evidence. Unknown is not zero. Ratings expire after 90 days; missing values do not enter averages or company scores. This supersedes earlier numeric ratings unless re-assessed under the current rubric.

    2026-10-22

    evidence

    Unattributed valuations, cumulative funding and headcounts are withdrawn, with originals archived.

    Dated financial observations remain separate from live market prices. OpenAI and Anthropic funding-round values now use first-party sources and the private valuation field; a round is not total funding. No such figures influence product or company scores.

    2026-10-22

    catalog

    Added eight products and three companies; availability does not automatically earn an adoption rating.

    Tinker, Inkling, Marble, Timestral, Beam, Muse Spark Open, Hermes Agent and Antigravity are cataloged with deployment trade-offs. Thinking Machines Lab, World Labs and Nous Research have externally available products. World Labs’ acquisition by AMD remains pending. Sierra and Retool modules are consolidated into their suites.

    2026-10-22

    release

    Solar retirement, MI400 launch and model-host availability change the deployment decision.

    Solar Pro 2/3 end service October 30 KST. AMD has launched MI400/Helios, but the hardware comparison keeps measured MI355X results. OpenAI models are available on Bedrock; Google Cloud general availability is planned for the coming weeks. Mistral Large 4 is an API preview with weights planned for late October.

    2026-10-22

    rating

    Customer evidence, inference economics and improving baselines now constrain Protected and Breakout.

    Generic platform/proprietary keywords, model ownership and specialization no longer earn app/model moat bonuses. Customer deployment, retention and structural advantages need dated sources. Comparative quality/cost/latency evidence expires after 90 days. Undisclosed margins remain unknown; Impact still measures adoption.

    2026-10-22

    catalog

    Company product portfolios
    Impact-normalized percentages and dollar allocations
    Owner-matched products and assessed evidence

    Removed unsupported product shares of company value throughout the company directory.

    The previous calculation divided company value by product Impact and assumed a 100% or 15% AI allocation. Neither assumption established revenue or valuation shares. Canonical product ownership now drives portfolios; unassessed rows do not contribute product Impact to company scores.

    2026-10-22

    rating

    gpt-oss
    Impact 84
    Not assessed

    Open weights and benchmark results do not establish adoption or a share of OpenAI's value.

    gpt-oss is a separately deployable open-weight model family. Recorded evidence supports availability, but not product-specific active deployment counts, revenue, or the former 8.4% financial attribution.

    2026-10-22

    evidence

    Updated ownership, public-parent classification, brand, Grok portfolio and Cursor acquisition.

    The xAI route remains stable. SpaceX's October 20 SEC disclosure confirms Cursor's August 14 acquisition. The AI division is not separately listed; transaction values and segment revenue are not current standalone market cap or Grok product revenue.

    2026-10-22

    catalog

    Separate buyer choices, cloud access links and refreshed Chinese model families; Jev, Seedance and MiMo added.

    GPT and Claude versions remain within their model families. Codex, Claude Code, OpenCode and Pi are harnesses; web/desktop/mobile editions remain within app rows. Cloud hosts are access routes, not duplicate models. Dates record a scoped review, not a recertification of the entire catalog.

    2026-10-22

    rating

    Furiosa RNGD / NXT RNGD Server
    Impact 78
    Impact 74 · named deployment

    Named regional production evidence supports the tier-1 hardware ceiling.

    The September Upstage case records Solar Pro 3 API and Daum search serving alongside H200. Chip specifications, access locations and scaled power comparisons do not establish worldwide adoption or single-server MLPerf throughput.

    2026-10-22

    rating

    GPT-Rosalind
    Impact 79
    Not assessed

    Specialized scientific positioning does not establish adoption or a durable model advantage.

    Trusted access and scientific workflows are documented, but product deployment volume and a current named-baseline advantage remain unestablished. Model, ChatGPT Workbench surface and API billing are distinguished.

    2026-09-29

    rating

    A.
    Impact 84
    Impact 80

    Narrower call-and-meeting utility, not a default general-purpose recommendation.

    Company-reported suite reach remains recorded, but feature-level engagement, paid demand and international adoption are not established.

    2026-09-29

    rating

    Hancom Assistant
    Impact 80
    Impact 76

    Office market share is not Assistant adoption.

    The product is offered, but Assistant-specific seats and revenue remain undisclosed.

    2026-09-29

    catalog

    Gemini Enterprise Agent Platform
    Six overlapping rows
    One platform

    Vertex AI and suite modules consolidated into their successor platform.

    Google says Vertex AI services continue under Agent Platform. The separate employee-facing Gemini Enterprise app is not merged.

    2026-09-29

    evidence

    Veo
    Impact 84 · unsupported 10M users
    Impact 82 · users undisclosed

    Old API endpoints retired; Veo 3.1 remains available.

    The Gemini API lifecycle table distinguishes retired Veo 2/3.0 endpoints from current Veo 3.1. Bundled Google distribution is not Veo adoption.

    2026-09-29

    catalog

    Higgsfield
    Impact 84

    Added the missing creative suite and company profile.

    Company-reported use and commercial scale support inclusion. Unspecified users and annualized revenue are clearly qualified.

    2026-09-29

    catalog

    Removed the protocol from commercial product rankings; learning resources retained.

    MCP is an interoperability standard. Its foundation is not a commercial vendor profile supported by this catalog.

    2026-09-29

    rating

    Llama
    Impact 84
    Impact 80

    Historical ecosystem influence no longer described as current dominance.

    Released weights remain available. Removed unsupported leadership claims and Behemoth from downloadable variants.

    2026-09-29

    rating

    Stable Diffusion
    Impact 84 · 10M users
    Impact 80 · users undisclosed

    Available self-hosted option, not an assumed current market leader.

    Current Stable Diffusion availability is verified; 10M users and ecosystem dominance were unsupported.

    2026-09-29

    rating

    Replit
    Impact 84 · 30M users
    Impact 80 · AI users undisclosed

    Removed legacy platform registrations from AI Agent adoption.

    Current releases confirm the service is active. Funding, revenue targets and old registrations are not AI Agent active users.

    2026-09-02

    listing

    Unitree Robotics
    Private
    Public · SSE 688836

    Listing status changed after its STAR Market debut.

    The company began trading on August 19, 2026; the profile now uses public-company classification and a dated market-cap observation.

    2026-09-02

    listing

    Mech-Mind Robotics
    Public · HKEX 9615

    Added after its Hong Kong listing with a verified commercial product suite.

    Company eligibility is supported by reachable first-party Mech-Vision, Mech-Viz, and Mech-Station product surfaces.

    2026-09-02

    listing

    SEER Robotics
    Public · HKEX 6106

    Added as a public embodied-intelligence infrastructure company.

    The Robot Brain platform is externally sold and the company reports 2,500+ company customers worldwide; that portfolio count remains qualitative rather than product userCount.

    2026-09-02

    release

    Tencent Hy
    Hunyuan / Hy3
    Tencent Hy4 Preview

    Current generation moved to Hy4 Preview; the product family was updated from the former Hunyuan branding.

    Generations remain in the versions block so the canonical buyer-facing family URL does not change every release.

    2026-09-02

    release

    GLM
    GLM-5.2
    GLM-5.3-Flash

    Current generation moved to GLM-5.3-Flash.

    The official model page documents native multimodal input, a 1M-token context window, and the current API model code.

    2026-07-29

    rating

    SuperLawyer
    Impact 84
    Impact 78

    Impact was reduced under the domain-blind adoption rule.

    Deep penetration among Korean lawyers supports Protected, but 6K users cannot carry the same Impact as horizontal products serving millions.

    2026-07-29

    rating

    Kimi
    Impact 88
    Impact 90

    Impact was raised after K3 established a global open-weight frontier role.

    K3's scale, open weights, and documented international benchmark standing meet the catalog's frontier-ecosystem tier, while undisclosed adoption prevents a higher score.

    2026-07-29

    catalog

    Product granularity

    Version, module, modality, and packaging rows were consolidated into buyer-facing product families.

    One durable page per brand or series concentrates evidence, links, comparisons, and search authority while versions remain auditable inside the row.