Download this edition as PDF
We'll email a 6-digit access code. Enter it to unlock the Daily Market Scan PDF.
Friday 21 August 2026
AI Daily Market Pulse · Weekly Digest
Three authorities asked for proof this week. None of them asked about the model.
Frontier and Industry Intelligence for regulated sectors: financial services, healthcare, energy and manufacturing. Edition date 2026-08-21, research window 2 to 21 August 2026, America/New_York.
The week in one paragraph
Between Monday 17 and Friday 21 August 2026, three separate authorities demanded evidence, and none of them asked how good anyone's model was. The FDA opened a docket on generative AI-enabled medical devices and asked 26 questions, most of them about what happens after clearance VERIFIED C02. Pennsylvania made grid and community commitments a precondition of a data-centre permit rather than a negotiating position VERIFIED C03. Texas kept roughly 250 to 300 interconnection requests on hold pending a project-by-project audit VERIFIED C04. In the same five days, Anthropic reported its first profitable quarter CITED C01, Cerebras launched a rack-scale system aimed squarely at inference economics VERIFIED C11, and CISA put an AI compute framework in the Known Exploited Vulnerabilities catalog with a three-day federal remediation clock VERIFIED C16. Capability and capital kept accelerating. The binding constraints moved to electricity, siting, and the ability to show your work after deployment.
What changed, in practical terms
For most of the last two years the gating question in a regulated enterprise was can the model do the task. This week that question was answered in three places by institutions that did not care about it. A permit officer in Harrisburg, a grid engineer in Austin and a reviewer in Silver Spring are all now asking a different question: can you evidence the thing you already built. That is a document-production problem and a telemetry problem, not a model-selection problem, and most agent programmes were not designed to answer it.
1. The three proof demands
All three landed inside a fifteen-day window. Read together they describe a single shift: the burden of proof moved downstream of the model, into siting, operations and postmarket behaviour.
Healthcare: the FDA asked about drift, not accuracy
The discussion paper published 18 August 2026 carries docket number FDA-2026-N-7874, poses 26 questions and proposes a two-axis risk framework to stratify generative AI-enabled devices. Comments close 19 October 2026. The agency states plainly that the paper does not propose or implement policy change VERIFIED C02.
Why it matters. A device whose output distribution can move after clearance breaks the assumption underneath traditional premarket review. The questions cluster around postmarket monitoring, which is the part almost nobody instrumented.
Energy: Pennsylvania priced the externality into the permit
Executive Order 2026-05, signed 18 August 2026, directs the Department of Environmental Protection to evaluate a data-centre permit only where the developer has made a legally binding commitment to the GRID requirements and has secured local approval. It removes AI data-centre proposals from Fast Track permitting and prohibits non-disclosure agreements. The Governor's office cited more than 100 proposals under discussion, most of them speculative VERIFIED C03.
Why it matters. Bring your own generation, fund your own infrastructure, and do it in public. That is a different capital stack than the one most 2025-vintage site plans assumed.
Energy: Texas is auditing the queue rather than growing it
Governor Abbott's 3 August 2026 directive had PUCT and ERCOT pause approvals of pending data centres while each project is verified. ERCOT is weighing more than 474 GW of proposed load, more than five times its record peak, with data centres roughly 90 percent of requests; grid officials confirmed the audit covers about 250 to 300 projects. ERCOT said it would miss the 7 August Batch Zero study deadline and would seek a good-cause exception VERIFIED C04. Wood Mackenzie analysis reported 12 August 2026 estimates US operators will commit to roughly 28 percent of the 1,066 GW requested nationally, leaving about 768 GW speculative or duplicative CITED C05.
The connective tissue. Texas and Pennsylvania are solving the same problem from opposite ends. Texas is filtering a queue that is mostly optionality. Pennsylvania is raising the entry price so the queue never fills with optionality in the first place. Either way, a compute plan that assumes interconnection is a scheduling problem is now a risk item.
2. Frontier ledger
Equal editorial weight, not equal praise. Company-reported figures are labelled. Benchmark claims are the vendor's own unless stated otherwise.
| Actor | Move | Date | Read |
|---|---|---|---|
| Anthropic | First profitable quarter reported: approximately USD 11.5 billion Q2 revenue and approximately USD 559 million positive adjusted operating income. Bloomberg reports run-rate above USD 65 billion. CITED C01 | 17 Aug | Company-reported and press-reported. The measure is adjusted operating income, and at least one published critique disputes the profitability framing on that basis. Report the measure, not the headline. |
| Anthropic | Machine-readable marking in Claude outputs for products released from 2 August 2026, for EU AI Act Article 50(2). Claude Code auto mode became the default for Pro, Max and Team from 14 August. CITED C15 | 2 and 14 Aug | A vendor shipping a compliance primitive and a default-autonomy change in the same fortnight. The second one is the change your control owners need to know about. |
| OpenAI | Presence, an enterprise platform for realtime voice and chat agents that query systems, take approved actions and escalate. GPT-5.6 adds multi-agent orchestration and programmatic tool calling. Press reports describe preparation for a share sale above USD 1 trillion against roughly USD 25 billion revenue and an estimated USD 14 billion 2026 loss. CITED C08 | August | The listing and loss figures are press-reported, not company-confirmed. Presence is the substantive item: approval and escalation are now product surface, not integration work. |
| xAI | Grok 4.6 released with a 500,000-token context window, configurable reasoning effort and Amazon Bedrock availability. Grok Bot, described as always-on AI teammates with a cloud computer per account, entered beta 11 August. CITED C07 | 11 and 12 Aug | Company-reported specifications. A persistent per-account compute environment is an identity and egress question before it is a productivity question. |
| SpaceX and Cursor | SpaceX completed the all-stock acquisition of Anysphere, developer of Cursor, valued at approximately USD 60 billion, via approximately 391 million new Class A shares, confirmed by SEC Form 8-K. VERIFIED C06 | 14 Aug | Largest startup acquisition on record as reported. Coding-agent infrastructure is now owned inside a defence-adjacent private company. Supplier concentration and jurisdiction belong in your vendor file. |
| Demis Hassabis stepped back from day-to-day Google DeepMind operations on 5 August, remaining Chair and Alphabet Chief Scientist; Koray Kavukcuoglu took the frontier lead; Jeff Dean and three senior researchers departed. Alphabet states nearly 90 percent of Fortune 100 companies use Gemini Enterprise. CITED C09 | 5 Aug | Adoption figure is company-reported. Leadership churn at the research layer rarely moves a roadmap inside two quarters, but it does move key-person risk in a vendor assessment. | |
| Meta | Muse Code shipped 5 August on Muse Spark 1.2, company-reported 82.9 percent on Terminal-Bench 2.1 against Claude Opus 5 at 86.7 percent. Contributor tier at roughly USD 0.10 and USD 0.20 per million tokens trades price for training rights over prompts and completions. CITED C10 | 5 Aug | The pricing tier is the story. A regulated buyer cannot take the contributor rate without a data-classification answer, and the discount is large enough that someone in engineering will try. |
| Cerebras | CS-4 announced, rack-scale on three new Wafer Scale Engines, first of the Nexus architecture. Company claims up to twice CS-3 speed, up to 30 times GPU tokens-per-second-per-user and up to 10 times CS-3 throughput per watt. VERIFIED C11 | 18 Aug | Performance figures are company-reported and not independently benchmarked here. Throughput per watt is the number to watch given the interconnection constraints above. |
| Alibaba and Moonshot | Qwen3.8 Max released 2 August and Qwen3.8-27B on 14 August; Moonshot released Kimi K3 at 2.8 trillion parameters in July. Twelve models from seven providers shipped in August to date. CITED C12 | 2 to 14 Aug | Open-weight capability keeps arriving on a monthly cadence. For regulated buyers this is a jurisdiction and provenance question as much as a capability one. |
| Obsidian Security | USD 85 million Series D at a USD 1.1 billion valuation announced 4 August, company stating 60 of the Fortune 500 as customers, extending agent governance controls to Claude Code and Claude Cowork. CITED C25 | 4 Aug | Capital is flowing to agent visibility rather than agent capability. That is a useful signal about where buyers are actually stuck. |
3. The gap the surveys keep measuring
Two independent bodies of evidence pointed at the same hole this month, from opposite directions: one from executives describing their own readiness, one from researchers measuring whether agents behave the same way twice.
The research says finishing is not the same as being reliable
A Princeton-authored paper accepted to ICML 2026 evaluated 15 models across two benchmarks and found recent capability gains produced only small reliability improvements, proposing 12 metrics spanning consistency, robustness, predictability and safety. Separate analyses of 15 major agent benchmarks, including AgentBench, WebArena and SWE-bench, conclude that evaluation methodology rather than model capability is the primary bottleneck on reliable deployment CITED C19.
Read this next to the FDA docket. A regulator asking how you monitor an adaptive device postmarket, and a research community saying task completion is a poor proxy for reliability, are describing the same missing instrument. If your agent programme reports completion rates and cost per task and nothing else, you are measuring the thing that both audiences have just told you is insufficient.
4. Sector reads
Each sector gets a win, a constraint and one control you can put in place inside a quarter.
Financial services
Win
Supervisory clarity is arriving from the jurisdictions that move first. MAS confirmed in a written parliamentary reply on 5 August 2026 that autonomous AI agents sit inside its supervisory expectations, following the November 2025 consultation on Guidelines on AI Risk Management and the Project MindForge toolkit CITED C24.
Constraint
The revised US interagency model risk guidance issued 17 April 2026, replacing SR 11-7 after fifteen years, states that generative and agentic AI models are not within its scope. The agencies have signalled a request for information to follow VERIFIED C13.
Control
Do not wait for the RFI. Run agentic use cases through your existing model risk lifecycle voluntarily, and document the decision to do so. When scope arrives, an institution with two years of validation artefacts is in a materially different conversation than one starting from a blank template.
The live pressure point in banking and insurance right now is the customer-facing voice channel, where OpenAI's Presence and comparable platforms have made deployment a configuration exercise rather than an integration project CITED C08. That is precisely where the exclusion in the revised guidance bites hardest: a voice agent that takes an approved action on a customer account is doing something a model risk framework would recognise, inside a scope that currently says it does not apply VERIFIED C13.
Healthcare
Win
The FDA is building the framework in public and asking for input rather than issuing rules and litigating them afterwards. A two-axis risk stratification and a 26-question request for feedback is an unusually open door for device makers and health systems, open until 19 October 2026 VERIFIED C02.
Constraint
There is no comprehensive federal AI statute for health. Organisations sit under a patchwork of FDA, CMS and HHS oversight layered with a growing set of state health AI laws, and the discussion paper explicitly does not resolve whether its approaches fall within existing authorities VERIFIED C02.
Control
Stand up postmarket output monitoring now, before it is required: sample clinical outputs on a fixed cadence, store the model version and prompt scaffold against every sampled output, and set a drift threshold that triggers human review. That is also the substance of a credible comment on the docket.
Note the boundary carefully. Most generative deployments in health systems today are administrative rather than clinical, and platforms like OpenAI Presence are being pointed at scheduling, intake and benefits questions precisely because those sit outside device classification CITED C08. That boundary is thinner than it looks: an intake agent that summarises symptoms for a triage queue has started influencing clinical judgement, which is the line the FDA framework is drawn around VERIFIED C02.
Manufacturing
Win
Humanoid deployment has crossed from demonstration into measured shift work. Figure AI has 40 Figure 03 units at BMW Spartanburg after an eleven-month trial, with trade press reporting ten-hour shifts and more than 90,000 parts handled; Agility reports Digit past 65,000 operating hours across nine customer facilities CITED C20 CITED C21.
Constraint
The headline reliability numbers, including the reported above 99 percent placement accuracy, are company and trade-press reported and have not been independently audited CITED C20. Procurement teams are being asked to underwrite uptime commitments against figures nobody outside the vendor has verified.
Control
Write the acceptance test before the purchase order. Define the cell, the part mix, the shift length and the failure taxonomy, then measure for one full production quarter against your own baseline. Vendor hours are a marketing metric. Your hours on your parts are an operating metric.
Energy
Win
Operational patterns are maturing faster than the autonomy debate suggests. Argonne National Laboratory's GridMind combines specialised agents with power-system analysis as an operator co-pilot, and Idaho National Laboratory recommends phased deployment running decision support in shadow mode alongside human decisions before autonomy increases CITED C22.
Constraint
Load, not compute, is the limiter. Texas paused roughly 250 to 300 projects against more than 474 GW of requests VERIFIED C04, Pennsylvania made grid commitments a permit condition VERIFIED C03, and only about 28 percent of the 1,066 GW requested nationally is expected to convert CITED C05.
Control
Adopt shadow mode as the default first stage for every operational agent, with an explicit graduation test. Log the agent recommendation and the operator decision side by side for a defined period, and only widen autonomy where the two agree at a rate you set in advance. Air-gapped deployment remains the pattern for nuclear and critical control environments.
5. Security note the operators should not skip
CISA added CVE-2025-62593 to the Known Exploited Vulnerabilities catalog on 17 August 2026 with a federal civilian remediation deadline of 20 August. The flaw is in Ray, the distributed computing framework used widely for AI training and serving, affects versions prior to 2.52.0 and carries a CVSS 4.0 score of 9.4. It stems from insufficient protection of Ray dashboard and API endpoints against browser-based attack, exploitable through DNS rebinding, and reported exploitation includes a DDoS botnet and a cryptomining campaign targeting GPU clusters VERIFIED C16.
The practical point. This is AI infrastructure entering the KEV catalog, not an application bug. If your data science teams stood up Ray clusters outside the change-management path, that is now a known-exploited exposure with a federal clock attached. Ask the question this week.
6. Workforce
The labour data continues to say something narrower and more useful than the discourse. Aggregate displacement is not visible. Entry-level compression is.
BLS projection work associates each 10-percentage-point increase in AI exposure with roughly a 0.6-point reduction in projected employment growth. Workers aged 22 to 25 entering AI-exposed roles show a reported 14 percent decline in job-finding rates, and entry-level employment in the most exposed occupations declined about 6 percent from late 2022 to July 2025 while older cohorts grew 6 to 9 percent. Analyses from Anthropic, the IMF and Stanford find the unemployment gap between exposed and insulated workers small and statistically insignificant CITED C23.
The operating implication. The pressure point is the on-ramp, not the workforce. If your agent programme removes the tasks juniors used to learn on, you have not saved cost, you have deferred a capability bill to the point where your seniors retire. Deloitte's own finding that 75 percent of leaders think human collaboration with agents beats automation alone points the same direction CITED C17.
7. Practitioner desk
Six questions we were actually asked this week, answered the way we would answer them on a call.
Our agents are in production and legal is now asking about the EU. What actually applies today?
Article 50 transparency obligations have been applicable since 2 August 2026: disclose when a person is interacting with an AI system, mark synthetic content, and handle emotion recognition and deep fakes. Penalties reach EUR 15 million or 3 percent of worldwide annual turnover, and a transitional period runs to 2 December 2026 for marking and detection on generative systems already on the market. Annex III high-risk obligations are placed at December 2027 and Annex I at August 2028 under the Digital Omnibus. Several vendor write-ups claim high-risk applied from 2 August; primary legal sources say otherwise and we follow them FLAG CITED C14.
One practical consequence: if you deploy a customer-facing voice agent, the disclosure obligation is live now regardless of where high-risk classification lands. OpenAI's Presence platform ships escalation and approved-action surfaces that make that disclosure easier to evidence, but the obligation is yours, not the vendor's CITED C08.
Our coding assistant just changed its default to a more autonomous mode. Is that a control event?
Yes, and it is the kind that slips past change advisory boards because no purchase order moved. Claude Code sessions for Pro, Max and Team defaulted to auto mode from 14 August 2026 unless a user or administrator pinned another setting, replacing repeated approval prompts with a classifier screening tool calls CITED C15. Treat vendor default changes as configuration drift: subscribe to release notes, pin the setting at the administrator level where your policy requires it, and re-baseline quarterly.
How do we evaluate the cheap contributor pricing tier engineering is asking for?
Meta's contributor tier is roughly an order of magnitude below standard pricing in exchange for use of prompts and completions as training data CITED C10. The evaluation is not commercial, it is data classification. Run it as a three-question test: can any prompt in this workflow carry customer, patient or material non-public information; can you prove it cannot; and would you be comfortable explaining the arrangement to your regulator. If any answer is uncertain, the tier is not cheaper, it is unpriced risk.
Our board asked whether we should pause agent work until regulators finish.
The evidence points the other way. Financial services has explicit exclusion of agentic AI from model risk scope VERIFIED C13, healthcare has a discussion paper that expressly changes no policy VERIFIED C02. Neither is a moratorium; both are open windows. The institutions that spend that window building evidence, postmarket monitoring, decision logs, reversal procedures, will be describing a working system when the rules land. The ones that pause will be describing an intention.
Our data-centre partner says interconnection is on track. How do we test that?
Ask for three artefacts rather than a status. First, the queue position and study batch with dates. Second, evidence of firm generation, since Pennsylvania now requires developers to bring their own and fund the infrastructure serving their load VERIFIED C03. Third, local approval status, which in Pennsylvania is now a precondition rather than a later step. If the partner is in ERCOT, ask directly whether the project sits inside the audited 250 to 300 VERIFIED C04.
What single metric would you add to our agent programme this quarter?
Recommendation-versus-decision agreement rate under shadow mode, measured per workflow and trended weekly. It is the metric that Idaho National Laboratory's phased pattern is built around CITED C22, it directly answers the reliability critique from the evaluation research CITED C19, and it is the only number we know of that a regulator, an operator and a CFO all read the same way.
8. Take this with you
Friday artefact
The Reg-Ready Field Note: Proof of Power, Proof of Behaviour
A seven-page field note for regulated operators, published alongside this edition. It converts the week into two evidence packs you can assemble without a new platform: a siting and energy evidence pack for anyone with a compute footprint, and a postmarket behaviour evidence pack for anyone running agents in a supervised process. Includes a fourteen-point scoring sheet, a shadow-mode graduation test, and the four questions to put to a data-centre partner before signing.
If you only do one thing next week
Pick the single agent workflow closest to a regulated decision and answer four questions in writing: what did it decide, on what evidence, who could reverse it, and how long would reversal take. If any answer requires a conversation rather than a query, that is the gap the FDA, the Pennsylvania DEP and your own model risk function are all going to ask about, in that order.
Research base
25 claim groups. Chips read VERIFIED when the item is named, dated and publicly checkable, CITED when the source is named but not independently re-verified, and FLAG when the point is contested and pending re-verification.
- C01 Anthropic first profitable quarter, Q2 2026 revenue approximately USD 11.5 billion and positive adjusted operating income of approximately USD 559 million, reported 17 August 2026; Bloomberg reports annualized run-rate above USD 65 billion. Company-reported and press-reported; the profitability framing is disputed on the basis that the measure is adjusted. Forbes · Bloomberg · Dissenting analysis
- C02 FDA, Considerations for the Regulation of Generative AI-Enabled Medical Devices: Discussion Paper and Request for Feedback, 18 August 2026, docket FDA-2026-N-7874, 26 questions, two-axis risk framework, comments close 19 October 2026, no policy change proposed. FDA discussion paper · FDA press announcement · MD+DI
- C03 Pennsylvania Executive Order 2026-05, Protecting Pennsylvania Consumers from Data Center Impacts, signed 18 August 2026: binding GRID commitment and local approval as permit preconditions, removal from Fast Track permitting, prohibition on non-disclosure agreements, more than 100 proposals under discussion. Executive Order text · Commonwealth release · WHYY
- C04 Texas Governor Abbott directive of 3 August 2026 to PUCT and ERCOT pausing pending data-centre approvals for audit; more than 474 GW of proposed load, data centres approximately 90 percent of requests, roughly 250 to 300 projects audited; ERCOT sought a good-cause exception on the 7 August Batch Zero deadline. Utility Dive · Holland & Knight · Jones Day
- C05 Wood Mackenzie analysis reported by Bloomberg, 12 August 2026: roughly 28 percent of 1,066 GW of requested US data-centre interconnection capacity expected to be committed, leaving about 768 GW speculative or duplicative. Bloomberg · Energy Central
- C06 SpaceX completed the all-stock acquisition of Anysphere, developer of Cursor, on 14 August 2026 at approximately USD 60 billion via approximately 391 million new Class A shares, confirmed by SEC Form 8-K; reported as the largest startup acquisition on record. Yahoo Finance · Value Add VC
- C07 xAI released Grok 4.6 on 12 August 2026 with a 500,000-token context window, configurable reasoning effort and Amazon Bedrock availability; Grok Bot entered beta 11 August 2026. Company-reported specifications. xAI release notes · Evolink · Digital Applied
- C08 OpenAI Presence enterprise voice and chat agent platform, and GPT-5.6 with multi-agent orchestration and programmatic tool calling; press reports of preparation for a share sale above USD 1 trillion against roughly USD 25 billion revenue and an estimated USD 14 billion 2026 loss, not company-confirmed. OpenAI · VentureBeat · AI Weekly
- C09 Google DeepMind leadership change of 5 August 2026: Demis Hassabis steps back from day-to-day operations remaining Chair and Alphabet Chief Scientist, Koray Kavukcuoglu takes the frontier lead, Jeff Dean and three senior researchers depart; Alphabet states nearly 90 percent of Fortune 100 companies use Gemini Enterprise, company-reported. CNBC · Gemini Enterprise release notes
- C10 Meta Muse Code launched 5 August 2026 on Muse Spark 1.2, company-reported 82.9 percent Terminal-Bench 2.1 and 59.3 percent DeepSWE 1.1 against Claude Opus 5 at 86.7 percent; contributor tier at approximately USD 0.10 and USD 0.20 per million tokens in exchange for training rights over prompts and completions. CNBC · Product analysis
- C11 Cerebras CS-4 announced 18 August 2026, rack-scale on three new Wafer Scale Engines, first of the Nexus architecture; company claims up to twice CS-3 speed, up to 30 times GPU tokens-per-second-per-user and up to 10 times CS-3 throughput per watt. Company-reported performance. Cerebras investor release · Yahoo Finance
- C12 Alibaba Qwen3.8 Max released 2 August 2026 and Qwen3.8-27B on 14 August 2026; Moonshot AI Kimi K3 at 2.8 trillion parameters released 17 July 2026; twelve models from seven providers released in August 2026 to date. LLM Gateway timeline · AI Release Tracker
- C13 Federal Reserve, OCC and FDIC revised model risk management guidance issued 17 April 2026 replacing SR 11-7, stating that generative AI and agentic AI models are not within the scope of the guidance; a request for information addressing bank use of AI is planned. OCC news release · American Banker
- C14 EU AI Act Article 50 transparency obligations applicable from 2 August 2026, penalties to EUR 15 million or 3 percent of worldwide annual turnover, transitional period to 2 December 2026 for marking and detection on generative systems already on the market; Annex III high-risk placed at December 2027 and Annex I at August 2028 under the Digital Omnibus. Market commentary asserting high-risk applicability from 2 August 2026 is contested; primary legal sources followed. Article 50 text · European Commission FAQ · Morgan Lewis
- C15 Anthropic machine-readable marking in Claude outputs for products released from 2 August 2026 for Article 50(2) compliance; Claude Code auto mode default for Pro, Max and Team from 14 August 2026, replacing repeated approval prompts with a classifier screening tool calls. Anthropic newsroom · Artificial Lawyer · 9to5Mac
- C16 CISA added CVE-2025-62593, a Ray code-injection and remote-code-execution flaw affecting versions prior to 2.52.0 with CVSS 4.0 score 9.4, to the Known Exploited Vulnerabilities catalog on 17 August 2026 with a federal civilian remediation deadline of 20 August 2026; exploitation reported via DNS rebinding, including a DDoS botnet and GPU cryptomining campaign. CISA KEV catalog · The Hacker News · Security Affairs
- C17 Deloitte agentic AI research, August 2026, 501 senior manager to C-suite respondents in the United States across five industries fielded April to June 2026, all at least piloting agents: 74 percent expect nearly half of processes redesigned around agents within four years, 61 percent expect most agents generally autonomous with human oversight, 75 percent agree human collaboration with agents beats automation alone, 5 percent report processes highly prepared. Deloitte Insights · State of AI in the Enterprise
- C18 McKinsey 2026 AI Trust Maturity Survey: progress in trust maturity alongside persistent gaps in strategy, governance and risk management; nearly two-thirds of organisations have not begun scaling AI across the enterprise. McKinsey
- C19 Princeton-authored paper accepted to ICML 2026 evaluating 15 models across two benchmarks, finding small reliability gains from capability improvements and proposing 12 metrics across consistency, robustness, predictability and safety; separate analyses of 15 agent benchmarks identify evaluation methodology as the primary bottleneck; KDD Workshop on Evaluation and Trustworthiness of Agentic AI held 9 August 2026, Jeju. Standardized AI evaluation paper · KDD workshop · Springer review
- C20 Figure AI: 40 Figure 03 units at BMW Spartanburg following an eleven-month Figure 02 trial, trade press reporting ten-hour shifts, more than 90,000 parts handled and above 99 percent placement accuracy; BotQ producing approximately one unit per hour with the 1,000th reported 23 July 2026; BMW extending to Plant Leipzig. Company and trade-press reported, not independently audited. Technology.org · Humanoid Guide
- C21 Agility Robotics reports Digit past 65,000 operating hours across nine customer facilities including GXO, Schaeffler, Toyota Motor Manufacturing Canada and Mercado Libre. Company-reported. Technology.org · Solid Market Research
- C22 Argonne National Laboratory GridMind agentic operator co-pilot combining specialised agents with power-system analysis; Idaho National Laboratory phased pilot approach running AI decision support in shadow mode alongside human decisions before increasing autonomy; air-gapped deployment identified for nuclear utilities and critical grid control centres. Grid co-pilot analysis · POWER Magazine
- C23 Labour market evidence: BLS projection work associating each 10-percentage-point rise in AI exposure with roughly 0.6 points lower projected employment growth; reported 14 percent decline in job-finding rates for workers aged 22 to 25 in AI-exposed roles; entry-level employment in most-exposed occupations down about 6 percent from late 2022 to July 2025 against 6 to 9 percent growth for older cohorts; Anthropic, IMF and Stanford analyses finding the exposed-versus-insulated unemployment gap small. Anthropic Economic Index · ICLE evidence review
- C24 Monetary Authority of Singapore written reply to a parliamentary question on agentic AI in financial services, 5 August 2026, confirming autonomous AI agents fall inside MAS supervisory expectations; follows the November 2025 consultation on Guidelines on AI Risk Management and the Project MindForge toolkit. MAS · Baker McKenzie
- C25 Obsidian Security USD 85 million Series D at a USD 1.1 billion valuation announced 4 August 2026, led by Crescent Cove Advisors, company stating 60 of the Fortune 500 as customers and extending agent governance controls to Claude Code and Claude Cowork. Obsidian Security · SecurityWeek