Skip to content
NewsroomTechnology

July 16, 2026, 6:39 PM · Policy Impact Report · 10 min read

EU opens 11 Android features to AI rivals. Reliability remains unsolved

On 16 July 2026, the European Commission ordered Google to give rival AI services equally effective access to 11 Android features, with most changes due in Android 18 by 1 August 2027. The order removes a system-access disadvantage, but a controlled mobile benchmark in which three frontier models completed 47.8% to 55.6% of GUI-only tasks shows why equal access cannot guarantee equal performance. https://digital-markets-act.ec.europa.eu/developer-portal/interoperability/alphabet-specification-proceedings-interoperability-ai-services_en; https://ec.europa.eu/competition/digital_markets_act/cases/202629/DMA_100220_2683.pdf; https://aclanthology.org/2026.acl-long.278.pdf

By Cumulant Research

Hover or tap an underlined term to see its definition.

Rear view of a dark Google Pixel smartphone.
A Google Pixel handset, part of the Android device family covered by the EU's new interoperability requirements. Photo: Simon Waldherr, CC BY-SA 4.0, via Wikimedia Commons

The quick version

  • The binding decision covers 11 Android features spanning assistant activation, device context, actions inside apps, system controls, on-device models and background work. https://ec.europa.eu/competition/digital_markets_act/cases/202629/DMA_100220_2683.pdf
  • Most measures must arrive with Android 18 and no later than 1 August 2027. One provision allowing several assistants to listen concurrently for their own wake words has an Android 19 deadline of 1 August 2028. https://ec.europa.eu/competition/digital_markets_act/cases/202629/DMA_100220_2683.pdf
  • The order removes unequal access to Android plumbing. It does not require OpenAI to use every feature, establish that ChatGPT will qualify for restricted features or make one assistant reason as well as another. https://ec.europa.eu/competition/digital_markets_act/cases/202629/DMA_100220_2683.pdf
  • In MobileWorld's shared research framework, Gemini-3-Pro completed 55.6% of GUI-only tasks, GPT-5 completed 54.0% and Claude-4.5-Sonnet completed 47.8%. These were API models in an emulator, not the consumer Gemini, ChatGPT or Claude apps. https://aclanthology.org/2026.acl-long.278.pdf
  • An earlier European Android remedy offers a warning, not a forecast: researchers estimated that a search-engine choice screen reduced Google's overall EEA mobile-search share by less than 1 percentage point. https://www.aeaweb.org/articles?id=10.1257%2Fmic.20230193

Figure

The final order reaches across the mobile task chain

The 11 named Android features grouped by what an assistant must do

Task stageFeatures in the final orderWhy access matters
InvokeLong-press contextual invocation; always-on wake-word detectionLets a user start a rival assistant through system-level access points
UnderstandCentral app-data access; context-aware intelligence; ambient dataProvides the screen, app and sensor information needed to interpret a request
ActStructured app integration; screen automation; system integrationLets the assistant change an app or phone state instead of only giving instructions
Keep workingSystem-level on-device models; third-party model implementation; background executionProvides local computing resources and supports work while another app is visible

The four rows are Cumulant Research groupings of the 11 features listed in the final measures. They are not equal weights or separate regulatory scores.

Source: https://ec.europa.eu/competition/digital_markets_act/cases/202629/DMA_100220_2683.pdf; https://digital-markets-act.ec.europa.eu/developer-portal/interoperability/alphabet-specification-proceedings-interoperability-ai-services_en · Final measures adopted 16 July 2026

Why it matters

The decision could reduce an operating-system advantage enjoyed by Google's Gemini and widen the route into Android for competing AI assistants such as ChatGPT. For technology companies and investors, however, commercial impact will depend on certification, developer adoption, product integration and task reliability rather than regulatory access alone. For users, the practical test is whether rival assistants can securely understand context and complete actions without introducing additional friction or errors.

The finding

Byline

Cumulant Research

The European Commission's decision removes a specific barrier: Google may no longer reserve important AndroidAndroidGoogle's operating-system platform for smartphones, tablets and other devices. https://digital-markets-act.ec.europa.eu/developer-portal/interoperability/alphabet-specification-proceedings-interoperability-ai-services_en system features for its own services or give rivals a meaningfully worse route to those features. The published measures name 11 features and require Google's interoperabilityinteroperabilityThe ability of one company's service to use and work with technical features controlled by another company. https://digital-markets-act.ec.europa.eu/developer-portal/interoperability/alphabet-specification-proceedings-interoperability-ai-services_en solutions to be equally effective in dimensions including setup, ease of use, data-transfer speed and energy consumption. https://ec.europa.eu/competition/digital_markets_act/cases/202629/DMA_100220_2683.pdf

That is a consequential change in the plumbing of mobile AI. It is not a finding that ChatGPT already matches Gemini on Android, that OpenAI will implement every available connection or that either product can reliably finish a long chain of actions. Access is the road; the assistant must still navigate it without taking a wrong turn. https://ec.europa.eu/competition/digital_markets_act/cases/202629/DMA_100220_2683.pdf; https://aclanthology.org/2026.acl-long.278.pdf

Figure

The final order reaches across the mobile task chain

The 11 named Android features grouped by what an assistant must do

Task stageFeatures in the final orderWhy access matters
InvokeLong-press contextual invocation; always-on wake-word detectionLets a user start a rival assistant through system-level access points
UnderstandCentral app-data access; context-aware intelligence; ambient dataProvides the screen, app and sensor information needed to interpret a request
ActStructured app integration; screen automation; system integrationLets the assistant change an app or phone state instead of only giving instructions
Keep workingSystem-level on-device models; third-party model implementation; background executionProvides local computing resources and supports work while another app is visible

The four rows are Cumulant Research groupings of the 11 features listed in the final measures. They are not equal weights or separate regulatory scores.

Source: https://ec.europa.eu/competition/digital_markets_act/cases/202629/DMA_100220_2683.pdf; https://digital-markets-act.ec.europa.eu/developer-portal/interoperability/alphabet-specification-proceedings-interoperability-ai-services_en · Final measures adopted 16 July 2026

What the order actually opens

The first stage is activation. Google must let third-party services use a long press on the home button or navigation handle and receive contextual information such as the current screenshot or foreground app. It must also provide access to always-on wake-word detection, including operation while the screen is locked or battery-saver mode is active. https://ec.europa.eu/competition/digital_markets_act/cases/202629/DMA_100220_2683.pdf

The second stage is understanding. Subject to user consent and other safeguards, covered services can obtain app data that applications have chosen to share, use device context and access ambient inputs such as sensor or screen information. Without context, an assistant asked to add this flight to my calendar may not know which flight the user means. https://ec.europa.eu/competition/digital_markets_act/cases/202629/DMA_100220_2683.pdf

The third stage is action. Structured integrations can expose deliberate app functions, while screen automationscreen automationTechnology that reads an application's visible interface and operates its buttons and fields in a way resembling a human user. https://ec.europa.eu/competition/digital_markets_act/cases/202629/DMA_100220_2683.pdf can let an assistant inspect and operate an application's visible interface in a separate virtual display. System integration covers functions such as changing certain settings, controlling media, opening applications, managing alarms and interacting with messages. https://ec.europa.eu/competition/digital_markets_act/cases/202629/DMA_100220_2683.pdf

The final stage is resources. The order covers access to system-level on-device models, the implementation of third-party on-device models and background executionbackground executionThe ability of an application to continue approved work while it is not visible on the screen. https://ec.europa.eu/competition/digital_markets_act/cases/202629/DMA_100220_2683.pdf supported by processors and memory. These provisions matter because a task can fail if the operating systemoperating systemThe core software that manages a device and supplies common functions to applications. https://ec.europa.eu/competition/digital_markets_act/cases/202629/DMA_100220_2683.pdf pauses it when the user changes applications or turns off the screen. https://ec.europa.eu/competition/digital_markets_act/cases/202629/DMA_100220_2683.pdf

Equal access still comes with conditions

Equal access does not mean unrestricted access. The final measures classify centralized app-data access, context-aware intelligence, structured app integration, screen automation and some system integration as restricted features. Google may require an assistant to satisfy eligibility conditions covering functionality, user intent, model risks, privacy, security and developer practices. https://ec.europa.eu/competition/digital_markets_act/cases/202629/DMA_100220_2683.pdf

Certification cannot be entirely controlled by Google. The order requires a free program for trusted outside certification authorities and also requires Google to offer its own free certification route. Conditions must be transparent, proportionate and non-discriminatory, and Google cannot demand a higher integrity standard from rivals than it applies to itself. https://ec.europa.eu/competition/digital_markets_act/cases/202629/DMA_100220_2683.pdf

Users retain control. Google may use consent prompts, access indicators and revocation controls, provided those measures apply equally to Google's services and third parties. The order also allows strictly necessary and proportionate protections for the integrity of the operating system and device. https://ec.europa.eu/competition/digital_markets_act/cases/202629/DMA_100220_2683.pdf

The reliability problem appears after access

MobileWorld offers a controlled look at what can happen after an assistant receives the ability to see and operate applications. The ACL 2026 paper describes a benchmarkbenchmarkA fixed collection of tasks and scoring rules used to compare systems under controlled conditions. https://aclanthology.org/2026.acl-long.278/ of 201 tasks across 20 applications, including ordinary interface tasks, tasks requiring clarification from a user and tasks combining interface actions with external tools. https://aclanthology.org/2026.acl-long.278/

For the comparison most relevant to the Android order, the researchers placed GPT-5, Gemini-3-Pro and Claude-4.5-Sonnet inside a planner-executor frameworkplanner-executor frameworkA system in which one component decides the next action and another component performs it on the device. https://aclanthology.org/2026.acl-long.278.pdf and used UI-Ins-7B as the grounding executorgrounding executorThe component that converts an instruction such as tap the Send button into an action at the correct place on the screen. https://aclanthology.org/2026.acl-long.278.pdf. Think of the frontier model as the person giving directions and the grounding executor as the hand locating the correct button. A good plan can still fail if the hand taps the wrong place, and accurate tapping cannot rescue a bad plan. https://aclanthology.org/2026.acl-long.278.pdf

Figure

With the same research framework, all three models failed many mobile tasks

Reported success on MobileWorld's GUI-only task category using UI-Ins-7B as the grounding executor

Gemini-3-Pro
55.6
GPT-5
54
Claude-4.5-Sonnet
47.8

The proprietary models were accessed through their official APIs and tested with a planner-executor framework in an Android emulator. These results do not measure the consumer Gemini, ChatGPT or Claude applications. The paper reports point rates without confidence intervals for this comparison.

Source: https://aclanthology.org/2026.acl-long.278.pdf · task-success rate, percent · ACL 2026 paper

The reported GUIGUIA graphical user interface, meaning the visible buttons, menus and fields through which a person operates an application. https://aclanthology.org/2026.acl-long.278.pdf-only success rates were close: 55.6% for Gemini-3-Pro, 54.0% for GPT-5 and 47.8% for Claude-4.5-Sonnet. The 1.6-percentage-point gap between the first two values is descriptive, not evidence that one consumer application is superior. The paper does not report confidence intervals for these three rates, and the systems were tested through APIs in a research framework rather than through the companies' consumer apps. https://aclanthology.org/2026.acl-long.278.pdf

The benchmark used an Android emulatoremulatorSoftware that imitates a phone on another computer so researchers can run repeatable tests. https://aclanthology.org/2026.acl-long.278.pdf and a binary score: a task counted only when its objective was fully achieved. The authors note that emulators can differ from physical phones in rendering delay, touch response and hardware behavior, while binary scoring ignores partial progress. Those choices improve repeatability but narrow what the results establish. https://aclanthology.org/2026.acl-long.278.pdf

Why this is not a ChatGPT-versus-Gemini scorecard

GPT-5 in MobileWorld is not the ChatGPT Android application, and Gemini-3-Pro in the benchmark is not the consumer Gemini application. The models were called through official APIs and connected to the same research machinery. Consumer products can differ in prompts, memory, safety checks, app integrations, model routing, permissions and software updates. https://aclanthology.org/2026.acl-long.278.pdf

The EU decision does not name ChatGPT as a guaranteed beneficiary. It creates a route for third-party services that meet the relevant requirements. Reuters identified OpenAI as a likely beneficiary, but the final measure does not establish which features OpenAI will adopt or how its consumer application will perform after adoption. https://www.investing.com/news/stock-market-news/google-required-to-open-up-to-ai-search-engine-rivals-under-eumandated-changes-4795675; https://ec.europa.eu/competition/digital_markets_act/cases/202629/DMA_100220_2683.pdf

The defensible conclusion is narrower. Google-controlled access was one source of potential advantage for Gemini. The order is designed to remove that source of unequal treatment. Any remaining performance gap could come from the model's planning, the software that translates plans into taps, product design, permission choices, certification, app support, device hardware or implementation quality. https://ec.europa.eu/competition/digital_markets_act/cases/202629/DMA_100220_2683.pdf; https://aclanthology.org/2026.acl-long.278.pdf

An older Android remedy shows why design matters

The closest historical comparison is imperfect but useful. Google introduced a search-engine choice screen on new Android devices shipped into the European Economic AreaEuropean Economic AreaThe European Union together with Iceland, Liechtenstein and Norway. https://www.efta.int/eea and the United Kingdom following the Commission's 2018 Android decision. The version studied by Francesco Decarolis, Muxin Li and Filippo Paternollo initially appeared during setup of new devices rather than across the installed base. https://www.android.com/choicescreen/; https://www.aeaweb.org/articles?id=10.1257%2Fmic.20230193

Figure

An earlier Android remedy moved EEA search share only slightly

Observed values, a model-implied selection rate and a causal estimate are kept separate

QuantityResultMeaning
Pre-remedy Google share98.67%Average EEA mobile-search share from April 2016 through February 2020
Google selection rate95.21%Model-implied choice-screen rate under the authors' device-stock assumptions
EEA market-share effectLess than 1 percentage pointPublished causal estimate for overall mobile search

The 95.21% figure was inferred using assumptions about the stock and replacement of devices. It was not a direct count of choice-screen selections. This older search intervention is evidence that remedy design and user behavior matter, not a forecast for AI assistants.

Source: https://www.aeaweb.org/articles?id=10.1257%2Fmic.20230193; https://igier.unibocconi.eu/sites/default/files/media/attach/AEJMicro-2023-0193.R2.pdf?VersionId=CxvbZvgwy9RvyXb5Dp75nOxqamoaKcJI; https://www.android.com/choicescreen/ · Study published August 2025; principal analysis through January 2022

The researchers used difference-in-differencesdifference-in-differencesA method that estimates an intervention's effect by comparing the change in an affected group with the change in an unaffected comparison group. https://igier.unibocconi.eu/sites/default/files/media/attach/AEJMicro-2023-0193.R2.pdf?VersionId=CxvbZvgwy9RvyXb5Dp75nOxqamoaKcJI, comparing the change in affected EEA countries with changes in comparison countries. They estimated that the remedy reduced Google's overall EEA mobile-search share by less than 1 percentage pointpercentage pointThe arithmetic gap between two percentages, so a movement from 20% to 21% is a change of 1 percentage point. https://www.aeaweb.org/articles?id=10.1257%2Fmic.20230193. They also calculated a 95.21% Google selection rate, but that figure was implied by a model using assumptions about device activation, retirement and accumulation rather than directly observed selections. https://www.aeaweb.org/articles?id=10.1257%2Fmic.20230193; https://igier.unibocconi.eu/sites/default/files/media/attach/AEJMicro-2023-0193.R2.pdf?VersionId=CxvbZvgwy9RvyXb5Dp75nOxqamoaKcJI

That study does not predict the effect of the new AI order. Search engines and assistants are different products, and a setup choice screen is different from opening operating-system capabilities. Its relevant lesson is limited: removing a formal restriction does not determine adoption or competitive impact when rollout design, product quality and user behavior still differ. https://www.aeaweb.org/articles?id=10.1257%2Fmic.20230193

The test begins with Android 18

Figure

Equal Android access is a staged rollout, not an immediate product change

The Android and search-data decisions have separate milestones

  1. 2026-01-27

    Proceedings opened

    The Commission opened separate Android-interoperability and search-data proceedings.

  2. 2026-04-27

    Android draft published

    The Commission published proposed measures and invited comments.

  3. 2026-07-16

    Binding decisions adopted

    The Commission finalized the Android and search-data measures.

  4. 2027-01

    Search-data pricing milestone

    Google must finalize its pricing offer under the separate search-data decision.

  5. 2027-08-01

    Most Android measures due

    The final Android text ties the measures to Android 18 and sets this outside deadline.

  6. 2028-08-01

    Concurrent wake words due

    Android 19 must allow several services to use always-on wake-word detection concurrently, within device limits.

Most Android obligations are tied to Android 18 and have an outside deadline of 1 August 2027. Concurrent wake-word operation for several services is due with Android 19 and no later than 1 August 2028.

Source: https://digital-strategy.ec.europa.eu/en/news/commission-opens-proceedings-assist-google-complying-interoperability-and-online-search-data; https://ec.europa.eu/competition/digital_markets_act/cases/202629/DMA_100220_2683.pdf; https://digital-markets-act.ec.europa.eu/developer-portal/data-access/alphabet-specification-proceedings-sharing-google-search-data_en

Most of the Android measures must be implemented in Android 18 and no later than 1 August 2027. The exception highlighted in the final text is concurrent always-on wake-word detection for several services, which is tied to Android 19 and an outside deadline of 1 August 2028. Reuters' same-day account summarized the user-facing rollout as beginning with the next Android iteration, but the final published measures provide the more precise deadlines. https://ec.europa.eu/competition/digital_markets_act/cases/202629/DMA_100220_2683.pdf; https://www.investing.com/news/stock-market-news/google-required-to-open-up-to-ai-search-engine-rivals-under-eumandated-changes-4795675

The rules extend beyond Google's own phones. Alphabet must take the technical, contractual and legal steps needed for compliance on covered Android devices supplied by third-party manufacturers when the relevant feature exists and the device still receives specified Android or Google system updates. Actual coverage can therefore vary with the feature and update status of each device. https://ec.europa.eu/competition/digital_markets_act/cases/202629/DMA_100220_2683.pdf

What to measure after rollout

A credible evaluation should preregister common tasks, run rival assistants on the same supported phones and applications, record full completion rather than polished replies, separate permission denials from reasoning failures, and publish uncertainty around the resulting success rates.

Access is necessary, not sufficient

The barrier removed is discriminatory access to Android's system-level capabilities. Once the order is implemented, a qualifying rival should not lose merely because Google reserved the wake wordwake wordA spoken phrase that activates an assistant while the device is listening for it. https://ec.europa.eu/competition/digital_markets_act/cases/202629/DMA_100220_2683.pdf, contextual datacontextual dataInformation about the user's current situation on the device, such as the visible screen, open application or relevant stored data. https://ec.europa.eu/competition/digital_markets_act/cases/202629/DMA_100220_2683.pdf, app-control channel or background resources for its own services. https://ec.europa.eu/competition/digital_markets_act/cases/202629/DMA_100220_2683.pdf

What could still keep ChatGPT or another rival behind is everything that happens after access: whether the provider builds the integration, obtains user permission and certification, supports the relevant devices and applications, plans the correct sequence and verifies that the task truly finished. MobileWorld's reported rates show that even strong models sharing the same research framework often failed controlled tasks. They do not prove how the consumer apps will rank after Android 18. https://aclanthology.org/2026.acl-long.278.pdf; https://ec.europa.eu/competition/digital_markets_act/cases/202629/DMA_100220_2683.pdf

Market reaction versus economic effect

This article makes no claim about Alphabet's share-price reaction. It also makes no claim that the order has already changed competition, adoption or consumer welfare because the principal Android obligations have not yet reached their implementation deadline. https://ec.europa.eu/competition/digital_markets_act/cases/202629/DMA_100220_2683.pdf

What to watch

  • How Google implements the measures in Android 18 and whether rivals receive access that is equally effective in practice.
  • Which AI providers qualify for restricted features and which Android capabilities they choose to adopt.
  • Whether certification, consent prompts or application support create meaningful friction for third-party assistants.
  • Whether real-world task-completion rates improve as models, interface-control systems and Android integrations mature.

How we did this

  • We used the European Commission's final published Android measures, rather than the April draft, to identify the 11 covered features, safeguards and implementation deadlines. https://ec.europa.eu/competition/digital_markets_act/cases/202629/DMA_100220_2683.pdf
  • We grouped the 11 features into four task stages, invocation, understanding, action and continued work, without assigning equal weight or claiming the groups are the Commission's own taxonomy. https://ec.europa.eu/competition/digital_markets_act/cases/202629/DMA_100220_2683.pdf
  • We extracted MobileWorld's GUI-only rates from the paper's main results table and checked its description of the shared planner-executor framework, UI-Ins-7B grounding executor, official API access, emulator and binary scoring. https://aclanthology.org/2026.acl-long.278.pdf
  • We did not convert MobileWorld's reported percentages into completed-task counts. The paper's task-summary table lists 116 GUI-only tasks, while the published percentages and later appendix accounting are not straightforwardly reconcilable with a single binary run, so the chart reproduces the authors' rates exactly. https://aclanthology.org/2026.acl-long.278.pdf
  • We treated the 2025 search-choice-screen study as historical evidence about remedy design, not as a forecast for AI assistants. Observed shares, model-implied selections and the authors' causal estimate are labeled separately. https://www.aeaweb.org/articles?id=10.1257%2Fmic.20230193; https://igier.unibocconi.eu/sites/default/files/media/attach/AEJMicro-2023-0193.R2.pdf?VersionId=CxvbZvgwy9RvyXb5Dp75nOxqamoaKcJI
  • The reporting cutoff was 16 July 2026. We separated the Android order from the parallel search-data order and used the Commission's primary pages for each timeline. https://digital-markets-act.ec.europa.eu/developer-portal/interoperability/alphabet-specification-proceedings-interoperability-ai-services_en; https://digital-markets-act.ec.europa.eu/developer-portal/data-access/alphabet-specification-proceedings-sharing-google-search-data_en

What this cannot establish

  • The Commission has published the final measures but says the full non-confidential decision explaining its reasoning will be published later. https://ec.europa.eu/competition/digital_markets_act/cases/202629/DMA_100220_2683.pdf
  • No source reviewed establishes that OpenAI has committed to implement all 11 features or that the ChatGPT application has completed certification for restricted features.
  • MobileWorld evaluates API models inside a research framework and Android emulator, not the consumer ChatGPT, Gemini or Claude applications. https://aclanthology.org/2026.acl-long.278.pdf
  • The MobileWorld paper does not report confidence intervals for the three GUI-only percentages, and its published task-count details are not fully reconcilable with converting those percentages into exact completed-task counts. https://aclanthology.org/2026.acl-long.278.pdf
  • Most Android obligations are not due until Android 18 and 1 August 2027, so no post-implementation comparison of rival consumer assistants was available at the reporting cutoff. https://ec.europa.eu/competition/digital_markets_act/cases/202629/DMA_100220_2683.pdf
  • The earlier choice-screen study concerns mobile search and a different remedy. It cannot establish the likely competitive effect of opening Android features to AI assistants. https://www.aeaweb.org/articles?id=10.1257%2Fmic.20230193

This is AI-assisted analysis under stated assumptions; it is not investment advice or a price target. Figures are as of the publication date and trace to the cited sources; markets and disclosures change.

Sources

  1. 01Decision of 16 July 2026, Final Measures, DMA.100220, European CommissionPrimary
  2. 02Alphabet specification proceedings, Interoperability for AI services, European CommissionPrimary
  3. 03Commission provides guidance to Google for AI interoperability on Android and sharing of Google Search data, European CommissionPrimary
  4. 04Commission opens proceedings to assist Google in complying with interoperability and online search data sharing obligations, European CommissionPrimary
  5. 05Consultation on interoperability measures for Google Android under the DMA, European CommissionPrimary
  6. 06Case Summary, Google Android interoperability, European CommissionPrimary
  7. 07Annex, Draft Google Android interoperability measures, European CommissionPrimary
  8. 08Alphabet specification proceedings, Sharing of Google Search data, European CommissionPrimary
  9. 09Google required to open up to AI, search engine rivals under EU-mandated changes, ReutersSecondary
  10. 10EU forces Google to share search data and open Android to rival AI companies, Associated PressSecondary
  11. 11MobileWorld: Benchmarking Autonomous Mobile Agents in Agent-User Interactive and MCP-Augmented Environments, Association for Computational LinguisticsAcademic
  12. 12MobileWorld paper PDF, Association for Computational LinguisticsAcademic
  13. 13Competition and Defaults in Online Search, American Economic AssociationAcademic
  14. 14Competition and Defaults in Online Search, accepted manuscript, IGIER, Bocconi UniversityAcademic
  15. 15About the Android choice screen, GooglePrimary
Artificial intelligenceAndroidAlphabetOpenAIEuropean UnionDigital Markets ActCompetition policyData journalismAlphabetGoogleOpenAIAnthropicEuropean UnionEuropean Economic Area

Related

Tech

Kimi K3 was 48% cheaper than Opus 4.8 on one AI benchmark, but only 10% cheaper than GPT-5.6 Sol

Moonshot AI released Kimi K3 on July 16, 2026. Artificial Analysis reported a lower average token cost per Intelligence Index task, but the small gap with GPT-5.6 Sol could disappear if K3 produces even modestly fewer acceptable results in a real workflow.

Rows of black server racks with cables and indicator lights inside a data center.
Tech

The EU can now fine general-purpose AI providers. We put the one-year odds of a first decision at 35%

The European Commission's power to impose fines under Article 101 began applying on 2 August 2026, although the detailed procedural regulation takes effect on 10 August. After tracing the required legal steps and correcting the closest DSA precedent to 833 days, we estimate a 35% chance of a first fine decision by 2 August 2027. [AI Office FAQ](https://ai-act-service-desk.ec.europa.eu/en/faq?faq_category_id=69) [Implementing Regulation 2026/1755](https://eur-lex.europa.eu/eli/reg_impl/2026/1755/oj/eng)

Exterior of the Berlaymont building, headquarters of the European Commission in Brussels
Tech

Alphabet raised $74.4 billion outside its operations while trailing cash still covered capex

Alphabet recorded $49.6 billion of equity proceeds and $24.8 billion of debt proceeds in the second quarter as its quarterly free cash flow turned negative. The financing proves that outside capital has joined the funding mix, but trailing operating cash still exceeded capital spending by $53.3 billion, so the accounts do not show that borrowing or issuing shares was unavoidable.

Exterior of Google's data center complex in The Dalles, Oregon, beneath a blue sky
Tech

TSMC needs $17.1bn of second-half operating cash to cover a $64bn capex year

TSMC generated $46.914bn of operating cash in the first half of 2026. If cash capex reaches the $64bn ceiling and is measured on the same basis as TSMC's quarterly figures, another $17.086bn from operations would cover it before dividends, but that is a break-even threshold, not a forecast or a verdict on whether the investment will earn an adequate return. [TSMC financial statements](https://investor.tsmc.com/english/encrypt/files/encrypt_file/reports/2026-07/114aaca0fea2050e96b91fffbab9ed04ba09cd92/FS.pdf) [Focus Taiwan](https://focustaiwan.tw/business/202607160019)

Aerial view of the TSMC Fab 21 semiconductor plant under construction in Phoenix, Arizona.