Digital Colliers Daily Briefing — July 17, 2026
Three stories today capture how the AI industry's frontier is being contested on distinct fronts. Brussels handed down its most consequential Digital Markets Act specification decisions to date, forcing Google to open Android to rival AI assistants and share search data with competitors. In Beijing, Moonshot AI released Kimi K3, a 2.8-trillion-parameter open-weight model that vaults China back to the top of the open-model conversation. And in Ulsan, unionized Hyundai workers walked off the job over humanoid robot deployment plans — the first automotive strike specifically targeting humanoid automation.
1. Brussels forces Gemini off its Android pedestal and cracks open Google's search index
The European Commission on Thursday issued two binding DMA specification decisions that reshape how Google must operate in Europe. On Android, Google will be required to give third-party AI assistants the same system-level access Gemini enjoys today — including "Hey Google"-style voice activation, screen content access, and the ability to take actions inside apps. On search, Google must share the data it uses to optimize its own search index with rival search engines and AI chatbots, subject to anonymization. Google has until January 2027 to begin sharing search data and until July 2027 to open up Android, according to The Verge.
The Commission was blunt about its intent: as it told The Register, prior voluntary commitments were "ineffective so far." The specification decisions are not noncompliance findings — they are formal instructions telling Google exactly what compliance under its gatekeeper obligations must look like. That matters because it forecloses the ambiguity Google has previously used to slow-walk DMA obligations.
Why it matters. Gemini's current advantage on Android — preload, hotword access, deep OS integration — is a structural moat that competitors like OpenAI, Anthropic, and Perplexity cannot replicate through OEM deals alone. Removing that moat in the EU's ~450 million-user market is the most direct regulatory intervention yet into mobile AI distribution. The search-data provisions are arguably more consequential still: rival search engines and AI labs gain access, under audited conditions, to the very query data that has underwritten Google's index quality for two decades.
Who is affected. Google faces meaningful erosion of its most defensible distribution channel and its most valuable proprietary dataset. Rival AI assistant vendors — including OpenAI, Anthropic, Perplexity, and European players like Mistral — gain a viable path onto Android without OEM gatekeeping. Search challengers and AI chatbot builders gain training and ranking signal they could not otherwise obtain. Android OEMs lose some vetting authority, which Google's Kent Walker framed as a security risk. Users get choice; whether they exercise it is another matter.
Who is unhappy. Google's Kent Walker warned the decisions "risk undermining vital privacy and security guardrails," and the Google-funded Chamber of Progress invoked Cambridge Analytica as a cautionary analogy — though as The Register notes, the Commission's FAQ spells out user bundling in cohorts of at least 1,000, suppression of rare-item records, direct and indirect identifier removal, mandatory independent audits, and Google's own final say over what gets shared.
What to watch next. Google's right of defense and near-certain litigation will likely stretch enforcement well past the 2027 deadlines. Watch whether OpenAI or Anthropic move quickly to negotiate Android integrations that anticipate the July 2027 opening, and whether the EU's biennial review cycle tightens or relaxes data-sharing conditions based on early practical experience.
Sources:
2. Moonshot's Kimi K3 puts a 2.8-trillion-parameter open model at the frontier
Moonshot AI announced Kimi K3 on Thursday: a 2.8-trillion-parameter mixture-of-experts model with 1-million-token context, native multimodal input, and 16 of 896 experts active per token. It is available today on Kimi.com and the Kimi API, with full weights promised by July 27. If Moonshot delivers on that date, K3 will be the largest open-weight model ever released, roughly doubling the size of DeepSeek's 1.6T v4 Pro.
The architecture introduces Kimi Delta Attention, Attention Residuals, and a Stable LatentMoE routing framework, which Moonshot claims yields roughly 2.5× the scaling efficiency of Kimi K2. Pricing lands at $3 per million input tokens and $15 per million output, with a 90% cached-input discount — matching Anthropic's Sonnet tier and making K3 Moonshot's most expensive model to date.
Independent evaluations have been striking. Arena ranked K3 #1 on Frontend Code Arena at 1679 points, jumping from #18 with a 76% pairwise win rate — ahead of Claude Fable 5 (63%) and GPT-5.6 Sol (58%). Artificial Analysis placed it third overall on the AA Intelligence Index at 57, comparable to Opus 4.8 and GPT-5.5 but behind Fable 5 and GPT-5.6 Sol, at roughly half of Opus's cost per task ($0.94 vs. $1.80). Moonshot itself concedes a "noticeable gap in user experience" versus the top closed models.
Why it matters. For nine of the past twelve months, Moonshot's models have set the upper bound on open-model scale. K3 pushes that boundary from trillion-class into 3T-class territory and — critically — matches its capability claims with credible third-party benchmarks. As Simon Willison noted, K3 is significant, though he pointed readers toward benchmark caveats rather than leaderboard hype. Artificial Analysis flagged one real regression: hallucination rate worsened to 51% on AA-Omniscience despite accuracy gains.
Who is affected. Anthropic and OpenAI now face an open-weight competitor within touching distance of Opus 4.8 at Sonnet-tier pricing — margin pressure that will compound as agent stacks commoditize model choice. Enterprise buyers gain a credible non-US, non-closed alternative, though the paradox remains: Moonshot recommends deployment on supernode configurations of 64+ accelerators, meaning "open weights" does not translate to cheap self-hosting. The vLLM project got day-zero KDA prefix-caching support contributed directly by Moonshot, signaling this is not a marketing-only open release. As TechCrunch reported, Moonshot is said to be raising fresh capital at a $31.5 billion valuation, up from $20 billion in May.
What to watch next. Whether weights actually ship by July 27; third-party throughput and cost figures once independent inference providers spin up; K3's performance on hidden and contamination-resistant evals like LiveBench; and whether Moonshot closes the UX and post-training gap it publicly acknowledges. Early hands-on reports describe the model as verbose and slow — currently only shipping with "max" thinking effort — which will shape how it slots into agent harnesses.
Sources:
- [HN · 1605↑] Kimi K3: Open Frontier Intelligence — Hacker News
- [AINews] Kimi K3 2.8T-A50B: the largest open model ever released; Opus 4.8-class at Sonnet 5 pricing — Latent Space
- Kimi K3, and what we can still learn from the pelican benchmark — Simon Willison
- not much happened today — smol.ai News
- Moonshot's upcoming Kimi 3 is expected to close the gap with Anthropic's Opus 4.8 — TechCrunch AI
- China has a new top model — Platformer
3. Hyundai's Ulsan workers stage the auto industry's first humanoid-robot strike
Thousands of unionized workers at Hyundai's Ulsan production complex — the world's largest automotive plant — began partial strike action after 15 rounds of negotiations with the company broke down over humanoid robot deployment plans. Day and night shifts ended two hours early from July 13 through July 15, and workers plan four-hour strikes from July 20 through 22. The Wall Street Journal, cited by Ars Technica, calls it "the car industry's first factory stoppage addressing humanoid robots."
The trigger is Hyundai Motor Group's plan to deploy the latest Atlas humanoid — the two-legged, 6-foot-plus, 100-pound-lifting robot built by Boston Dynamics, which is about to become a wholly owned Hyundai subsidiary. Union objections began the moment Atlas was unveiled earlier this year.
Why it matters. Every prior industrial robotics wave — fixed-arm assembly automation, AGVs, collaborative robots — was eventually absorbed into labor contracts through attrition, retraining clauses, and productivity-sharing agreements. Humanoids are different in kind: they are designed to occupy human workspaces and perform human tasks without retooling the environment. That makes them substitutional in a way earlier automation was not, and unions are treating them accordingly. Ulsan is the first test case of what humanoid deployment negotiation actually looks like.
Who is affected. Hyundai loses production hours during a period when Boston Dynamics integration is meant to be a strategic advantage. Boston Dynamics faces its first sustained labor backlash tied to a specific product deployment. Other automakers — BMW, Mercedes, Ford, and Chinese OEMs piloting Figure, Apptronik, Unitree, and UBTech humanoids — will watch the Ulsan outcome closely for negotiating precedent. So will the humanoid vendors themselves, whose commercial thesis depends on frictionless factory-floor adoption.
What to watch next. Whether Hyundai's July 20-22 escalation forces a return to the table, and what the eventual settlement looks like — job guarantees, deployment caps, retraining provisions, or productivity-sharing formulas. Any framework agreed at Ulsan will become the reference template for humanoid negotiations across the global auto sector and eventually beyond it.
Sources:
Today's three stories describe the same underlying dynamic from three angles: the concentration of AI capability, and who gets to push back. Regulators are prying open the distribution and data moats that entrenched the current AI incumbents; an open-weight competitor is compressing the price of frontier intelligence itself; and organized labor is contesting the physical deployment layer where those models eventually touch work. The next twelve months will be defined less by model capability than by who negotiates its terms — Brussels, Beijing's open-weight labs, and unions on the factory floor are all now doing so simultaneously.