DEV Community: News

ESMA Final Report Proposes Simpler EU Taxonomy Reporting with a Narrower OpEx KPI

News — Fri, 30 Oct 2026 21:00:15 +0000

📅 29 October 2026 • European Securities and Markets Authority • ESMA Final Report Proposes Simpler EU Taxonomy Reporting with a Narrower OpEx KPI

On 29 October 2026, the European Securities and Markets Authority (ESMA) published its Final Report with technical advice to the European Commission on simplifying key performance indicators under the Taxonomy Disclosures Delegated Act (EU) 2021/2178. The document carries the reference ESMA32-846262651-5927 and follows a public consultation that ran from 1 July to 12 August 2026. ESMA has submitted the report to the Commission for consideration in its review of the Delegated Act.

A narrower OpEx KPI

The central proposal narrows the mandatory operating expenditure (OpEx) KPI to non-capitalised research and development expenditure, while the existing materiality reliefs remain in place. ESMA does not recommend introducing prescribed requirements for a voluntary OpEx+ KPI. Instead, it advises clarifying the status of voluntary Taxonomy-related disclosures and applying safeguards on reconciliation, transparency, comparability and greenwashing risks. The revised KPI framework would be monitored for five years.

Simpler group reporting

For group-level Taxonomy reporting, ESMA proposes removing the weighted-average group KPI and simplifying how the main reporting regime is determined, primarily by reference to the nature of the parent undertaking. Where groups have secondary businesses of a different nature, simplified disclosures are advised so that material information is not lost. The measures aim to make disclosures more proportionate, clearer and easier to use, while preserving the relevance, comparability and reliability of Taxonomy disclosures.

Coordinated action across the ESAs

The technical advice was coordinated with the other European Supervisory Authorities, in particular on the voluntary use of counterparties' OpEx information and group-level reporting. On the same day, 29 October 2026, the European Banking Authority (EBA) published its own Final Report advising on five areas, including removing the Fees and Commissions KPI and the Trading Book KPI for credit institutions and investment firms. Together, the two reports form a coordinated simplification agenda that now rests with the European Commission.

IMPACT

Reporting teams at non-financial undertakings and asset managers would need to redefine OpEx data collection around non-capitalised research and development expenditure.
Group finance functions would no longer calculate the weighted-average group KPI, reducing consolidation effort.
Any voluntary Taxonomy disclosure would require reconciliation and transparency safeguards.
The changes are technical advice, so application depends on the European Commission's review of and amendments to the Delegated Act.

Sources:

Continue the conversation: https://t.me/ComplianceExplained · Bluesky

SEC Issues Exemptive Relief from Treasury Clearing Requirements for Certain Inter-Affiliate and Non-U.S. Repo Transactions

News — Fri, 30 Oct 2026 20:59:58 +0000

📅 30 October 2026 • U.S. Securities and Exchange Commission • SEC Issues Exemptive Relief from Treasury Clearing Requirements for Certain Inter-Affiliate and Non-U.S. Repo Transactions

Targeted relief issued. On 30 October 2026 the U.S. Securities and Exchange Commission issued exemptive relief connected with its U.S. Treasury clearing rules in exemptive order Release No. 34-106789. The relief excludes from the definition of eligible secondary market transactions certain repo transactions, removing them from the scope that direct participants must submit for central clearing, subject to the limits and conditions stated in the order.

Two exclusions from mandatory clearing. First, certain repo transactions involving non-U.S. counterparties, allowed within specified limits and conditions. Second, certain repo transactions between a direct participant of a U.S. Treasury securities clearing agency and an affiliated entity that is not an investment company as defined in Section 3 of the Investment Company Act of 1940, provided the affiliate otherwise meets the definition of an affiliated counterparty in Rule 17ad-22(a).

Timeline of the rules. The Treasury clearing rules were adopted in December 2023. To allow further engagement on compliance, operational and interpretive questions and to support orderly implementation, the Commission extended the original compliance dates and confirmed its readiness to engage with market participants. In September 2025 the SEC launched a one-stop webpage consolidating implementation updates, FAQs, Commission and staff statements and other resources.

How the relief emerged. The SEC said the action follows extensive public engagement on the scope of the rules for these transaction types, including requests for relief from market participants published for comment. Commissioner Mark T. Uyeda, designated by Chairman Paul S. Atkins to lead the initiative overseeing the transition to central clearing of U.S. Treasury securities, said the relief addresses practical concerns while preserving the important benefits of central clearing.

Questions during implementation. Market participants preparing for compliance can contact SEC staff in the Division of Trading and Markets directly by email at tradingandmarkets@sec.gov or by phone at 202-551-5777. The relief is conditioned on the stated limits and definitions, so firms should verify eligibility transaction by transaction before adjusting clearing workflows.

IMPACT
• Direct participants of U.S. Treasury clearing agencies with non-U.S. counterparties or inter-affiliate repo.
• Global banking groups moving repo between affiliates, which may fall outside mandatory clearing.
• Compliance and operations teams mapping flows and verifying limits, conditions and affiliate status.
• Clearing agencies applying the exclusions in trade submission policies.
What changes: certain non-U.S. and inter-affiliate repo no longer requires central clearing submission if conditions are met. Risks: eligibility depends on definitions and limits.

Sources:

Continue the conversation: https://t.me/ComplianceExplained · Bluesky

Zest launches text-based AI dining concierge that votes and books reservations for groups

News — Fri, 30 Oct 2026 20:23:05 +0000

📅 October 30, 2026 • Source: techcrunch.com

Zest, a restaurant discovery app founded in November 2024, has introduced an AI dining concierge that lives in SMS. Users text the number (212) 470-6067, and the agent replies with suggestions built on the history of the user and their friends. The app launched in 2026, backed by the 1.8 million dollars in pre-seed funding raised at its founding from Alexis Ohanian of 776 and Steve Jang of Kindred Ventures. The concierge had been in testing since the previous month, and the company held the launch until the booking function was ready. CEO Mario Gomez-Hall described the shift in one line: “The agent is the product now.”

How the SMS concierge works
The agent draws on card spending data connected through Plaid, and Zest states it uses restaurant spending only. On that base it selects options by cuisine, occasion, dietary restrictions and neighborhood. In a one-to-one chat it can answer practical questions about a venue: how busy a restaurant is, whether it offers Wi-Fi, and it can send the menu. In a group chat the behavior changes: the agent blends the tastes of all participants, resolves indecision by running a vote directly in the thread, and books a table without leaving the conversation. Reservations pass through connected services, namely OpenTable, Resy and SevenRooms. On the business side, bookings and, soon, food orders are the first revenue lines, with the startup considering an expansion beyond restaurants to businesses and shopping.

IMPACT
For restaurants listed on OpenTable, Resy and SevenRooms, the service adds a text-based entry point where group decisions turn into confirmed reservations, placing the venue in front of diners at the moment of choice without a separate app install. Restaurant marketing teams can check that listing details such as menus, Wi-Fi and busy times are current in the reservation systems, since the agent reads them. For users, the data model has a stated boundary: Plaid connections cover restaurant spending only, so people enabling the feature can review what is shared. For consumer app teams, the launch is a case of distribution through a channel people already use, SMS, instead of a new app, and of holding a public launch until the transactional step works end to end.

TechCrunch reported the launch on October 30, the material is also available on tech.yahoo.com, and the launch page is on the company site. The concierge reaches users through a channel they already open daily, so the reservation flow starts where the group conversation already happens, and the assistant carries the context of that discussion into the booking request.

Sources: techcrunch.com and tech.yahoo.com

Continue the conversation: https://t.me/AIAutomationArchitecture · Bluesky

Chip startup PowerLattice raises $86M per SEC filing to ease AI power crunch

News — Fri, 30 Oct 2026 20:17:09 +0000

📅 October 30, 2026 • Source: geekwire.com

PowerLattice, a chip startup based in Vancouver, Washington, has raised 86 million dollars, according to a GeekWire report based on an SEC filing dated Friday, October 30. The filing shows the company is seeking up to 111 million dollars with 18 participating investors, while investor names and the round type remain undisclosed. The amount is nearly three times the 31 million dollars the company raised when it came out of stealth in November 2025. The news follows a technical milestone: earlier in the week, on October 26, the company demonstrated what it describes as the first monolithic integrated voltage regulator chiplet. Founded in 2023 by alumni of Qualcomm, NUVIA and Intel, the company has raised about 117 million dollars in total.

How the chiplet works
An integrated voltage regulator, or IVR, moves power conversion close to the chip that consumes it. In the monolithic design shown by PowerLattice, the controller, the power transistors and the inductors sit on a single piece of silicon rather than as separate components. In the published test, the chiplet powered a mockup of an AI accelerator load with sharp transients, the rapid swings that occur when an accelerator changes workload intensity. The company claims the approach reduces the power needs of AI compute by more than 50 percent and delivers approximately 2x performance at a fixed data center power budget. These figures are company statements from the demonstration, not independently measured results. According to the company, the first batch of chiplets is produced by TSMC together with a manufacturing partner whose name has not been disclosed. The earlier round, a 25 million dollar Series A, was led by Playground Global and Celesta Capital, and former Intel CEO Pat Gelsinger, a general partner at Playground Global, sits on the company board.

IMPACT
For data center operators, power delivery is a direct limit on how many accelerators a facility can run, so a supplier working on voltage regulation at package level addresses a line item present in every capacity plan. If the claimed reduction of more than 50 percent in power needs holds under independent measurement, operators could revisit rack density and cooling budgets, and approximately 2x performance at fixed power would change economics per accelerator. Procurement and vendor risk teams can read the filing as a runway signal: with about 117 million dollars raised and production at TSMC, the company is moving from demonstration toward manufacturing. Engineering teams evaluating IVR technology can watch for named partners, independent benchmarks and qualification data, and until those appear, the published figures remain vendor statements.

GeekWire reported the filing on October 30, and the demonstration is described in a press release on powerlatticeinc.com.

Sources: geekwire.com and powerlatticeinc.com

Continue the conversation: https://t.me/AIAutomationArchitecture · Bluesky

OpenAI to honor a shared robots.txt signal for AI training preferences across providers

News — Fri, 30 Oct 2026 20:16:54 +0000

📅 October 30, 2026 • Source: openai.com

OpenAI has announced that it will honor a shared signal in robots.txt based on the draft developed by the AI Preferences working group at the IETF. The announcement was published on October 30, 2026. The change targets a known friction point in web publishing: a site owner who wants to state preferences about AI training on their content currently writes separate rules for every crawler, one block per provider, and repeats the work each time a new crawler appears. Under the approach described by OpenAI, the owner declares those preferences once, and several participating AI providers read the same single signal. The company says the rollout will begin in the coming months and describes the initiative as voluntary, inviting other providers to support the same signal.

How the shared signal works
The signal is placed in robots.txt, the file site owners already use to manage automated access. Its syntax follows the draft developed by the IETF working group on AI Preferences, known as aipref. The practical difference is in the number of rules: instead of maintaining one set of directives per crawler, the owner expresses a training preference a single time, and the declaration applies across several providers that honor it. OpenAI states that its existing crawler controls remain in place and that the new signal does not replace them, so current robots.txt configurations keep working during the transition. Participation is voluntary on both sides: a site owner chooses whether to add the signal, and each AI provider chooses whether to honor it. OpenAI frames the announcement as an invitation to the industry and asks other providers to adopt the same signal, which would extend the reach of one declaration without extra work for the site owner.

IMPACT
For site owners and publishing teams, the immediate change is maintenance: one shared declaration can replace repeated per-crawler entries as participating providers switch on support in the coming months. Web and SEO teams can review their robots.txt now, decide what the training preference should say, and verify that existing per-crawler directives still apply, since OpenAI confirms both mechanisms coexist. For compliance and data governance functions, the signal adds a documented, machine-readable expression of preferences that can be referenced in internal policies on content use. The voluntary design carries a practical limitation: coverage depends on each provider adopting the signal, so a declaration does not by itself bind providers outside the initiative, and per-crawler rules stay relevant for them. During the rollout window, some providers may honor the shared signal while others do not, which argues for keeping both mechanisms in place until adoption is broad.

The announcement is available on openai.com, and the working group documents are tracked on ietf.org.

Sources: openai.com and ietf.org

Continue the conversation: https://t.me/AIAutomationArchitecture · Bluesky

Google's first orbital AI data center prototype runs Gemma 4 on a TPU in space

News — Fri, 30 Oct 2026 18:06:36 +0000

📅 October 30, 2026. Source: TechCrunch.

Google's Project Suncatcher has reached its first operating milestone: the prototype satellite is online and communicating with ground stations after launching earlier in October aboard a SpaceX Falcon 9 rocket together with more than 100 other spacecraft. TechCrunch reported the news on October 30.

An AI data center in orbit

The satellite carries a TPU, Google's custom AI accelerator, and that TPU is running a Google DeepMind Gemma 4 model in orbit. To verify that the hardware and the model behave correctly in space conditions, Google tested the setup by having the spacecraft compose a space-themed haiku and then limericks and report on orbital conditions. The experiment is an early proof of concept for Project Suncatcher, Google's effort to explore whether data center workloads can eventually run on satellites rather than only in terrestrial facilities.

Why orbit

The long-term idea behind the project is that satellites in a suitable orbit can receive abundant solar energy, which could in principle power compute hardware at scales that are difficult to reach on the ground. A single prototype does not settle the question, but it moves the concept from design documents to testable hardware. The parameters that matter now include how AI accelerators tolerate radiation, how thermal management works without ground-style cooling, and how much bandwidth the spacecraft can maintain with ground stations. TechCrunch notes that the news broke in the same week when Alphabet disclosed that AI features are hurting search ad revenue growth, a detail mentioned here for context only.

IMPACT

For infrastructure and aerospace engineers, an operating TPU in orbit makes Project Suncatcher measurable, and the coming months of telemetry will say more than the demo outputs: radiation resilience of AI accelerators, thermal behavior in vacuum and ground-link bandwidth are the figures that will determine feasibility. Organizations watching compute economics should treat orbital data centers as a long-horizon option rather than a near-term source of capacity, because one prototype sits years away from any scalable deployment. Compliance and legal teams in space, telecom and data governance roles may also note that orbital compute raises questions of licensing, spectrum use and data sovereignty that do not arise in terrestrial facilities. For now the practical takeaway is narrow: the experiment works, poetry included, and the hard engineering data is still ahead.

Power remains the catch: with roughly 1 kW of onboard generation the satellite can run the TPU only about 20 minutes at a time, and radiation tolerance in real flight is still unverified.

Sources: techcrunch.com, blog.google

Continue the conversation: https://t.me/AIAutomationArchitecture · Bluesky

PsiQuantum shows 8x8 optical switch for AI data centers with 1.2 dB loss

News — Fri, 30 Oct 2026 18:01:00 +0000

📅 October 30, 2026. Source: Tom's Hardware.

PsiQuantum, a company better known for photonic quantum computing, has unveiled an all-optical circuit switch aimed at AI data centers. In company-reported lab tests the 8x8 prototype typically loses only 1.2 dB fiber to fiber, before connector losses of light from fiber to fiber, compared with the 3 dB loss that data-center optics are normally designed to tolerate, and it achieves this without built-in amplification. Tom's Hardware covered the announcement on October 30.

How it works

The device passes light between fibers with no electrical conversion at all. It is built on etched silicon circuits and contains no moving parts, unlike the mechanical designs used in some rival switches. With no moving parts, the design avoids the mechanical trade-offs of MEMS-based alternatives. Optical circuit switching matters for AI clusters because it lets operators reconfigure which groups of GPUs communicate with each other without converting signals to electricity and back, reducing the power, amplification and latency that conventional re-cabling would require. In large training environments, network topologies often need to change as workloads change, and an optical switch can redraw those connections quickly.

Where it stands

The current prototype has 8 inputs and 8 outputs, which places it well behind competitors on port count. Google's MEMS-based Palomar switch reaches 136 ports, Lumentum offers 300, and Salience Labs has demonstrated 32. PsiQuantum says larger switches are in development, so the 8x8 device should be read as a technology demonstration rather than a shipping product. The low loss figure is nonetheless notable, because signal loss directly determines how far optical links can run and how much amplification a network needs.

IMPACT

For data-center and network engineers, the result addresses a genuine constraint: every decibel of loss saved reduces the amplification, power and cabling needed to rewire GPU clusters, which translates into lower operating costs and simpler physical design. If PsiQuantum can scale the port count while keeping the 1.2 dB figure, operators of large AI clusters would gain another option for reconfiguring topologies between training runs. Teams making infrastructure decisions today should treat this as early-stage technology, since an 8-port lab prototype is at least several quarters away from production relevance, and established vendors already ship switches with far higher port counts. Procurement and compliance teams evaluating photonic components should also watch how silicon-photonics supply chains fit into existing export-control and dual-use screening practices as the category matures.

Sources: tomshardware.com, psiquantum.com

Continue the conversation: https://t.me/AIAutomationArchitecture · Bluesky

GitHub turns Copilot into a multi-agent workspace with new Projects feature

News — Fri, 30 Oct 2026 18:00:44 +0000

📅 October 30, 2026. Source: InfoWorld.

A workspace, not a chatbot

At its GitHub Universe conference, GitHub launched a redesigned Copilot application that the company now positions as a central workspace for AI-driven software development. Instead of a single chat window, the new app connects AI agents with repositories, developer tools and project context, so that work can be planned, delegated and tracked in one place. The redesign effectively turns Copilot from a conversational assistant into an orchestration layer for development work, and GitHub describes the move as a bet on multi-agent software development, where several specialized agents cooperate on the same codebase under human supervision.

What Copilot Projects adds

The centerpiece of the update is a new capability called Copilot Projects. Developers can assign agents goals that span multiple repositories at once, a coordination task that previously had to be handled manually across projects. Projects also lets teams attach context from outside the code: Slack conversations, design mockups and meeting notes can all be supplied to an agent as background material. Multiple agents can run in parallel on different parts of the same effort rather than waiting in a single queue. GitHub says the goal is to let agents work with the same material a human team would use, not just with source code.

How the work gets done

Under the hood, Copilot distributes work across parallel agents, tracks their progress and returns results to the developer. The typical output is a pull request that a human reviews before anything is merged, which keeps people in control of what lands in the codebase. Because progress is tracked centrally, developers can see where each agent stands without chasing status updates. GitHub presents this as the operational model for the next phase of AI-assisted engineering: agents handle drafting, testing and iteration, while people set goals and approve outcomes.

IMPACT

For engineering teams, the change means rethinking how tasks are written: goals for agents must be specific enough to span repositories and end in a reviewable pull request, so task decomposition and clear acceptance criteria become daily skills. Team leads will need new routines for reviewing parallel streams of agent output, since several sets of changes can arrive at once. Security and compliance specialists face a new surface to govern: agents that read Slack conversations, meeting notes and design files work with data well beyond source code, so access policies, audit trails and retention rules need to cover agent context as well. Organizations adopting the workspace should plan a staged rollout, starting with low-risk repositories, and should expect an adjustment period of weeks to months while review workflows stabilize.

GitHub lists Copilot Projects in public preview, and analysts quoted by InfoWorld flag parallel-agent conflicts, growing token bills and code review as the next bottleneck. Early users are advised to start with low-risk repositories.

Sources: infoworld.com, github.com

Continue the conversation: https://t.me/AIAutomationArchitecture · Bluesky

FTC Removes 19 Websites Impersonating the Agency and Secures More Than $80 Million in Redress Under the Impersonation Rule

News — Fri, 30 Oct 2026 16:02:02 +0000

📅 29 October 2026 • U.S. Federal Trade Commission • FTC Removes 19 Websites Impersonating the Agency and Secures More Than $80 Million in Redress Under the Impersonation Rule

The U.S. Federal Trade Commission removed 19 websites that impersonated the agency and its officials, and has completed 14 enforcement actions under the Impersonation Rule with more than $80 million in consumer redress since the rule took effect in 2024. Social media pages impersonating the FTC and its officials were also taken down.

Enforcement actions

Since early April 2025 the FTC has brought eight enforcement actions alleging violations of the rule. Targets included an online lead generation firm in the health insurance market, a marketing company selling healthcare plans, a business operating as a debt collector and a company running a sham substance use disorder treatment clinic.

Other cases covered debt relief, tax debt relief and student loan debt relief schemes. The most recent action targets a sham credit repair operation that took about $200 million from consumers, including military service members. In August 2026 a federal court temporarily halted that scheme and froze its assets at the request of the FTC.

Websites taken down

Warning letters to domain registrars and other service providers over the past year led registrars to take down 19 websites impersonating the FTC, among them FTC-Support.org, ftcfederal.org and federaltradecommission.net. The letters identified apparent violations of the rule by the operators of these sites.

Rulemaking

Companies and individuals that violate the Impersonation Rule may be required to refund affected consumers and pay civil penalties of up to $53,088 per violation. In September 2026 the FTC opened an ANPRM on whether to update the rule or create a new one addressing the role of platforms in promoting impersonation scams, including ad-optimization tools offered by social media, search engines and other digital marketplaces. Comments are due 30 November 2026.

IMPACT

Who is affected: advertisers, lead generation firms, debt and tax relief providers, credit repair services, and any business claiming affiliation with a government agency.
What changes in processes: review of claims about government or business affiliation, verification of the authority behind each claim, screening of advertising platforms for impersonation campaigns.
Deadlines: comments on the platform ad rule are due 30 November 2026.
Risks: civil penalties of up to $53,088 per violation, consumer refunds ordered by federal courts, enforcement under a possible updated rule.

Sources:

Continue the conversation: https://t.me/ComplianceExplained · Bluesky

FCA Consults on Financial Crime Guide Overhaul: AI, Cryptoassets, Sanctions and Proliferation Financing

News — Fri, 30 Oct 2026 16:00:47 +0000

📅 23 October 2026 • Financial Conduct Authority (UK) • FCA Consults on Financial Crime Guide Overhaul: AI, Cryptoassets, Sanctions and Proliferation Financing

Scope and status of the consultation

On 23 October 2026 the FCA opened consultation CP26/38 on updates to its Financial Crime Guide, seeking feedback until 24 December 2026. The Guide is non-binding guidance and imposes no new requirements. The proposals cover seven areas: fraud, cryptoassets, financial and trade sanctions, proliferation financing, data security, financial crime systems and controls, money laundering and terrorist financing.

Fraud, sanctions and proliferation financing

The fraud section reflects the failure to prevent fraud offence introduced by the Economic Crime and Corporate Transparency Act 2023, and strengthens expectations on fraud risk assessment, governance and management information. The sanctions and proliferation financing sections are updated for the evolved UK sanctions regime and add good and poor practice examples from supervisory work. Data security material on AI, synthetic identity fraud and digital asset risk is refreshed.

Cryptoassets, unhosted wallets and Banking as a Service

The cryptoassets section prepares firms for the new authorisation regime: from October 2027 firms carrying out certain cryptoasset activities in the UK will need to be FCA authorised under the Financial Services and Markets Act 2000 (Cryptoasset) Regulations 2026, while remaining subject to the Money Laundering Regulations, the Terrorism Act 2000 and the Proceeds of Crime Act 2002.

Draft section FCG 3.2.8 lists applying enhanced due diligence to transactions involving unhosted wallets as good practice, and draft FCG 3.2.5 calls it poor practice to use third-party technology without being able to evidence the basis of findings or scoring. The systems and controls section covers Banking as a Service delivery models, including outsourced, white labelled and embedded finance, and reinforces that regulated firms stay accountable across the delivery chain.

Separately, on 23 October 2026 the FCA announced restrictions on 6 unregulated Annex 1 lenders over anti-money laundering concerns. The restrictions were imposed at each firm's own request, in the period from 21 April to 15 October 2026. The FCA is also assessing responses to an information request sent to around 900 registered Annex 1 firms.

IMPACT

• Affected: all firms the FCA supervises for financial crime, including Annex 1 institutions and cryptoasset businesses.
• Process changes: fraud risk assessment, governance and management information, sanctions and proliferation financing controls, data security, documentation of third-party tools, accountability across Banking as a Service chains.
• Timeline: responses by 24 December 2026, final Guide text in a Policy Statement, cryptoasset authorisation regime from October 2027.
• Risk: the guidance is non-binding, but control gaps can attract supervisory attention and enforcement.

Sources:

Continue the conversation: https://t.me/ComplianceExplained · Bluesky

DoorDash has seen only several hundred orders from third-party AI agents

News — Fri, 30 Oct 2026 15:22:33 +0000

📅 30 October 2026 • techcrunch.com

DoorDash CEO Tony Xu said on the company's Q3 2026 earnings call that DoorDash has seen maybe several hundred orders in total from third-party AI agents such as Muse, Instinct and Fo. For comparison, DoorDash reported 993 million total orders in Q3 2026 alone, roughly 11 million a day.

Xu described agent-based ordering as a very dynamic area and said users still predominantly place orders through the DoorDash app. He said the company wants to learn more about agentic use cases and develop a healthy relationship with third-party agent makers, adding that there could be an opportunity to advertise within those agents in the future. The comment came in response to an analyst question about whether AI agents could take over commerce and disintermediate platforms like DoorDash.

DoorDash is meanwhile building its own agent surface: a text-to-order agent launched in September, a CLI agent introduced in July, the Ask DoorDash chatbot launched in June, and an MCP ordering connector for corporate orders. The wider market is splitting on whether outside agents get access at all. Amazon blocks third-party agents, while Walmart welcomes them as a way to win business from competitors.

IMPACT

For anyone building or integrating commerce agents, the number resets expectations that agentic ordering is already a volume channel. Several hundred orders against roughly 11 million orders a day means the current wave of third-party agents is a rounding error in platform economics, and it helps explain why platforms can afford to experiment rather than defend. It also shows a strategic split in how platforms treat outside agents, which suggests that access rules, rather than agent capability alone, may decide where agents can actually transact. Teams planning agent products should treat platform policy as a primary design constraint, and treat agent-driven order volume as early stage until a platform reports otherwise. The gap between hundreds of orders and millions of daily orders suggests that consumer behavior changes more slowly than the technology, even as the underlying capabilities improve rapidly. For investors and partners, it also signals that the near-term opportunity may be in enabling infrastructure and tools rather than in expecting agents to drive meaningful transaction volume on their own. The result also undercuts the framing that agents will quickly displace apps in consumer food delivery, and it gives platform teams a concrete number to plan against.

Sources: techcrunch.com — gate.com — ir.doordash.com

Continue the conversation: https://t.me/AIAutomationArchitecture · Bluesky

Epoch AI: the US received 72% of the world's AI compute since 2022

News — Fri, 30 Oct 2026 15:11:52 +0000

⚠️ Not fully verified

📅 30 October 2026 • epoch.ai

Epoch AI published its AI Chip Imports Explorer on 27 October, a dataset that estimates where the world's AI chips actually end up. Its headline finding is that of 27.4 million H100-equivalents of AI compute that arrived globally from 2022 through June 2026, an estimated 19.8 million, or 72%, went to the United States.

China, counting domestic production, ranks second at 2.6 million H100-equivalents, or 10%, ahead of the European Union at 1.4 million, or 5%, and Malaysia at 0.7 million, or 2%. No other country received more than 2%. India and Japan each received about 1.6%, followed by Australia, the United Kingdom and Indonesia.

The American share rose across the period. It accounted for about 60 to 63 percent of each year's new AI compute from 2022 to 2024, then 72% in 2025 and 78% in the first half of 2026. Cumulative AI chip import value exceeds $400 billion. The estimates are built from customs records covering 16 major destinations, with Chinese domestic production estimated separately. The methodology uses a 2021 baseline to separate AI-attributable chips from general server and GPU imports, then nets out re-exports so each chip counts once at its final destination. H100-equivalents normalize different GPU models, from A100s to H200s and B200s, into a comparable unit of compute capacity.

IMPACT

For AI teams and companies that source compute internationally, the dataset provides a single consistent baseline for where capacity is concentrated and how quickly that concentration is changing. It is relevant to supply planning, to data residency and jurisdiction discussions, and to anyone tracking how export control policy maps onto physical chip flows. The figures are estimates rather than a registry. They track arrivals recorded in customs data and include China's domestic output, so the numbers describe shipments rather than commissioned or operational capacity, and the method notes that some chips may still be in transit. The dataset also reveals a structural split. China received 18% of chip units but only 10% of compute, meaning the chips China gets deliver less computing power on average, which is consistent with export controls limiting access to the most advanced models. Estimates for countries outside the US carry wider uncertainty, with China's range spanning 1.9 to 4.2 million H100-equivalents under alternative assumptions. Epoch AI publishes the dataset together with its methodology documentation.

Sources: epoch.ai — pivotnews.ai — officechai.com

Continue the conversation: https://t.me/AIAutomationArchitecture · Bluesky