AI Weekly | Agents Get Badges Before Jobs, as AI Giants Lose Billions and Write Their Own Rules

Sources: Reuters / BNN Bloomberg, CBC, NVIDIA, ZDNET, Forkast, Xinhua, Cailian Press, China News Service, SiliconANGLE, The Korea Times, DW and others.

The AI industry accelerated in two directions this week. Agents moved into real production systems, and the machinery around them — identity, audit, revocation, liability — became a product sold by the seat. At the same time, the companies building those agents started writing their own laws. Nvidia used GTC Berlin to ship a full "safe agent" reference architecture; ServiceNow, Microsoft, Workday and Okta put four layers of control on the market in the space of two weeks. Anthropic's draft IPO prospectus showed a US$80.7 billion net loss for the first half, while the company was pitching Congress a safety bill it drafted itself. In China, "AI Plus" travelled from a Central Committee document to a nine-ministry manufacturing campaign. Compute was the most split story of all: South Korean chip exports jumped 222.5% in the first 20 days of October, while global chip stocks shed almost US$2 trillion in eight trading sessions.

1. Agents get badges first, value arrives later

On October 21 at GTC Berlin, Nvidia announced the OpenShell Reference Architecture, packaging four jobs into one enterprise stack: Rubin runs models, Vera CPUs run agent harnesses, BlueField DPUs run the Sentry security framework for out-of-band monitoring and in-silicon policy enforcement. The same day, ServiceNow and Nvidia's Arc became generally available, starting on Mac with Windows to follow. Two days earlier, Microsoft expanded Project Perception with Guardian Agents, and Workday's Agent Passport entered early access with Cisco AI Defense as its launch assessor. Inside two weeks, Okta (Agent SSO on September 22), ServiceNow (AI Passport on October 15) and Microsoft (October 19) turned four responsibilities into purchasable products: know which agents exist, authorise what they can do, contain them, and adapt protection as risk changes.

The numbers explain why this market heated up so fast. Forkast calculates that Microsoft bundles Guardian Agents and Dynamic Agent Risk into Microsoft 365 E7 at US$99 per user per month — the governance tax lands per seat, whether or not that employee uses an agent. Cisco's March 2026 survey found 85% of organisations experimenting with agents but only 5% in broad production. Okta's own survey found 96% of executives confident in their IAM for non-human identities, yet only 34% of organisations apply the same controls to their digital labour force as to human staff. Saviynt's 2026 CISO AI Risk Report found 92% of security chiefs lack full visibility into agent identities. The measurement gap is worse: the MEASURE function of the NIST AI Risk Management Framework — the part that would tell you whether an agent's work is any good — scores 0.6 out of 5.0 in Oleria's practitioner assessment. Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027, citing inadequate risk controls.

Why now? ZDNET's answer is blunt: a summer of agent-led breaches. Enterprises are not refusing to deploy agents; they lack a mechanism that can explain, after an incident, what the agent did, who authorised it, and whether the next action can be stopped. So the pitch moved from capability to provability. What Nvidia is really selling is the enforcement point: once ServiceNow's Control Tower revokes an agent's passport, BlueField must carry out that revocation at machine speed in silicon, so compromised software never gets the chance to override policy. That shifts security from the software layer into hardware — and conveniently turns governance into the next hardware refresh cycle. That is the most important commercial crack of the week.

The examples are concrete. When Arc's passport is revoked, OpenShell contains the current task in a sandbox for later human audit while other approved actions continue without a restart (ZDNET). Cisco's adversarial test results are mapped to OWASP Top 10, MITRE ATLAS and NIST frameworks, and Workday issues "security stamps" on that basis. Asymptote Labs raised a US$5.5 million seed round and open-sourced Beacon, which covers Claude Code, Codex, Cursor and more than 20 other harnesses — deliberately splitting after-the-fact recording from before-the-fact blocking into two products.

The call: within 12 months, agent identity becomes a standard IAM module, and procurement questionnaires shift from "what can it do" to "what can you show me when it goes wrong". But the measurement layer is still empty. Tools can see what an agent did; they cannot say why it chose that, let alone whether the decision was right. Until 2027, the human review queue — not model capability — remains the real bottleneck on agent scale.

Sources: nvidia.com | zdnet.com | cisco.com | theaiinsider.tech | forkast.news | forkast.news | unite.ai | pulse2.com

2. The frontier labs' balance sheet, and the laws they want to write

On October 21, Reuters obtained the draft prospectus for Anthropic's planned IPO: a net loss of US$80.7 billion for the first half, against US$13 billion a year earlier. Of that, US$72 billion was an accounting charge tied to the rising fair value of convertible bonds, not cash spending. The cash story is a different line: nearly US$9 billion on computing capacity and infrastructure in the first half, up from US$2.3 billion, against a commitment to spend at least US$518 billion over a decade on cloud, compute and infrastructure. Revenue rose 45% quarter on quarter to US$16.8 billion in the third quarter, from US$11.6 billion, but the sequential growth rate slowed noticeably. Reuters says the company expects to list in November, with some investors weighing a valuation around US$1.5 trillion.

A longer view: in 2025 Anthropic posted a US$42 billion net loss, or an operating loss of about US$8.1 billion on US$4.6 billion of revenue once the accounting charge is stripped out. The operating loss rate is still extreme, and the decade-long cloud commitment has mortgaged future cash flow. Meanwhile, the draft bill Politico obtained sets its own terms: agencies could only move to block a model if it poses a credible risk of killing 50 or more people, and even then only with a judge's permission. Third-party auditors could self-certify, and federal law would preempt California's SB 53, New York's RAISE Act and Illinois' SB 315. Critics note this is weaker than July's FRONTIER Act, which required licensed third-party evaluators.

Put the two together and the week makes sense. A company reporting an US$80.7 billion loss hands Congress a bill framed as public protection, whose most substantive clauses move the power to block models away from expert agencies and into courts, while building a wall in front of state law. This is not simple regulatory arbitrage; it is a change of position — from regulated to rule-writer. After the leak, Anthropic's head of federal affairs Anthony Cimino conceded the preemption language was "not precise enough" and promised revisions.

The real constraints are growing elsewhere. The FTC is using civil investigative demands under Section 5 to probe model providers. Colorado's ADMT rules, effective in 2027, mandate 30-day adverse-outcome disclosure and a 45-day independent human review, with the reviewer holding override authority and the automated system barred from assisting. Taiwan, after indicting people involved in AI server smuggling, plans to amend the Foreign Trade Act to control AI hardware. The EU's Digital Omnibus adds human oversight requirements while deferring deadlines. India's CERT-In became the first regulator to treat agents as an active attack surface, recommending default-deny egress and least-privilege identities. Five jurisdictions, ten days, not one statute — agent liability is being assembled from subpoenas, rulemakings, criminal amendments and compliance deferrals.

The case file was dense this week. On October 15, eight states sued OpenAI in their own courts — Iowa, Kansas, Nebraska, West Virginia, Arkansas, Idaho and Montana with Republican attorneys general, Arizona with a Democrat — alleging violations of state consumer protection law. Iowa's attorney general Brenna Bird said she would not accept a big company making money by hurting Iowa's kids. On October 6, Postmedia, Torstar, The Globe and Mail and The Canadian Press sued Anthropic in Ontario, alleging it scraped their journalism to train Claude and circumvented technical protection measures. On October 20, Chapman University's annual Survey of American Fears put "AI harming the political process" at 51.9%, tied for second with another world war, with 49.7% worried about AI hurting the economy — the first time AI has entered the survey's top ten. For contrast, World Bank Managing Director Zhang Wencai told the Bund Summit that only 11% of emerging-market financial regulators use AI-enabled supervisory tools, and about half say they lack the staff and tools to supervise at all.

The call: Anthropic's November IPO is the pricing anchor for the entire AI trade this quarter; how the book builds will decide whether "compute for growth" can keep raising money. The bill will likely be watered down before it reaches committee. The binding constraints will come from litigation, state law and agency subpoenas first. Sam Altman's three endgames — an IAEA-style global regulator, a CERN-style joint lab, or one winner that "melts everyone else's GPUs" — got a clear ranking from him: the third is "obviously bad". Whether the market will pay for the first is a question this week answered with silence.

Sources: bnnbloomberg.ca | techbuzz.ai | lavx.hu | thepaper.cn

Sources: cbc.ca | sina.com.cn | pymnts.com | popsci.com

3. China's "AI Plus": from document language to factory language

On October 21, Xinhua published the full text of the Central Committee and State Council opinion on integrating technological and industrial innovation. In the section on upgrading traditional industries with advanced technology sits one line: fully implement the "AI Plus" initiative and enable AI across every sector. Five days earlier, on October 16, the State Council's executive meeting devoted a session to the progress of "AI Plus", calling AI "the key variable shaping future development" and stressing application-led innovation, coordination between computing and power, and industry-specific, tiered regulation. On October 19, the Commerce Ministry and eight other departments issued an opinion on technology-driven commerce that turns "AI Plus manufacturing" into a dedicated campaign, naming smartphones, PCs, home appliances, wearables and robots as the products to enrich.

The most informative part of this policy wave is the quantified targets. The seven-ministry 15th Five-Year Plan for advertising sets industry revenue growth of about 10% a year, notes that internet advertising already accounts for 89.1% of media advertising revenue, and explicitly requires regulating AI-generated advertising while encouraging industry agents and lightweight model tools. Locally, the 2026 Physical AI and World Models Conference ran in Nanjing from October 19 to 22, organised by Jiangsu Data Group, Nanjing Big Data Group and the Chinese Association for Artificial Intelligence — the country's first high-level conference dedicated to the field. Jiangsu's case is written in its supply chain: Suzhou alone has more than 160,000 industrial enterprises, with Suzhou Industrial Park on autonomous driving and compute chips, Xiangcheng on connected-vehicle testing, the high-tech zone on robot components, and Kunshan on industrial vision. PAIWorld, a world model built by the Chinese Academy of Sciences' Institute of Industrial AI, was shown at the event.

The language has changed from two years ago. Instead of "breaking bottlenecks", policy now treats scenarios as a resource to be allocated. The State Council talks about coordinating compute and power because the marginal cost of agents has moved from chips to electricity and data. It talks about tiered regulation because a single safety standard would both fail to contain frontier models and suffocate low-risk uses such as factory inspection. Industry moved in step: on October 20, DeepSeek open-sourced its internal inference framework DeepInfer, with distributed inference, speculative decoding, SWA Bounded Replay and structured output, compatible with both Nvidia and Huawei Ascend, plus a production token codec aligned with its own API. MiniMax's M3.1 Flash cut agentic task costs by 70%, runs inference nearly three times faster at a peak of 200 tokens per second, carries 428 billion parameters and a one-million-token context, prices at 15 cents per million input tokens, and scored 61.7 on OSWorld V2.

The harder line is silicon. DeepSeek's partnership with Huawei has moved past the supplier relationship: Huawei engineers are stationed at DeepSeek in phases, feeding real training requirements — how compute tasks are allocated, how processors exchange data — directly into chip designs due early next year. DeepSeek plans to begin pretraining on Ascend 950DT as early as the end of this year, and Huawei intends to allocate most of its 960DT output in the first quarter of 2027 to DeepSeek. IDC data shows Nvidia's share of China's AI chip market fell to 34% in the first half of 2026, with Huawei at 28% — a six-point gap. On the software side, DeepSeek's open-source TileLang compiler runs the same operators on Ascend, striking at the migration moat that keeps CUDA customers in place.

The call: watch three verifiable signals over the next two quarters. Does compute-power coordination produce actual projects? Does large-scale pretraining on Ascend 950DT actually run? Do "AI Plus manufacturing" orders show up in listed company filings rather than press releases? If all three hold, China's AI story shifts from catching up on models to deployment density — which is much harder to sanction.

Sources: news.cn | sina.com.cn | cls.cn | chinanews.com.cn | jschina.com.cn | siliconangle.com | sina.com.cn | 21jingji.com

4. Compute's two faces: record exports, retreating valuations

On October 21, Korea Customs Service data showed exports of US$61.6 billion in October 1-20, more than double a year earlier and a record for the period. Chip exports jumped 222.5% to US$27.5 billion, or 44.7% of total shipments — up 16.2 percentage points in a year — with exports to China up 154.4% to US$16.6 billion and to the US up 154.9% to US$10.7 billion. The same day, Japan's Ministry of Finance reported September exports up 16.7% year on year, a 13th consecutive gain but below the 20.1% median forecast, with exports to the US up 22.9% and to China up 8.6%. On October 20, Reuters reported that ByteDance, Alibaba and Tencent had ordered more than 3 million RTX PRO 5500 cards from Nvidia, worth over US$35 billion at list price. Total orders from the three exceed 6 million units, while Nvidia has told partners that global output in 2027 will be at most 3 million — meaning customers may ultimately receive 5% to 10% of what they ordered.

Share prices told the opposite story. Before October 19, chip stocks suffered their worst stretch since March 2025: close to US$2 trillion of market value gone in eight trading sessions, with Intel down more than 14% in a week and Micron down 14.6%. As of October 15, the world's 20 largest AI and semiconductor stocks had lost more than US$1 trillion in a week; SK Hynix fell more than 10% in a day, Samsung Electronics 8.73%, Advantest nearly 14%, Tokyo Electron 6.45% and Infineon 5%. On October 19 came the rebound: Intel rose 4.2%, AMD 3.5%, Marvell 4.7%, Broadcom 3.5% and Nvidia 1.9%.

Demand and prices are not contradictory; they are talking about the numerator and the denominator. The numerator is real: chips rising 16 percentage points as a share of Korea's exports in a single month means AI is rewiring a country's trade structure, and Japan's 13-month streak rests on chipmaking equipment. The denominator is rates: US Treasury yields above 5% make long-dated capital spending more expensive to finance. Broadcom CEO Hock Tan said financing, more than chip or power shortages, may be the bigger obstacle to AI expansion — a remark worth more than most analyst notes because it comes from the buyer's side of the table. The supply mix is shifting too: Groq, now effectively owned by Nvidia, is dropping its SRAM-centric design for 3D-stacked DRAM, using TSV vertical interconnects to add capacity rather than chase HBM-style bandwidth. Inference economics — dollars per parameter stored, not raw speed — is becoming the centre of chip design.

The cases are pragmatic. Chinese buyers cannot get B300 and must seek case-by-case approval for H200, so the de-rated but compliant RTX PRO 5500 became the best available option at about US$11,400 a card, with samples in December and commercial delivery in the first quarter of 2027; Shenzhen-listed Tansun Technology said on October 13 it would buy 20,000 units, warning that shortages could delay delivery. Memory signals point the same way: CXMT announced a new architecture breakthrough over the weekend, Elon Musk said China will produce "a lot of memory chips" in 2028, and the market has begun to worry about a memory glut in 2029.

The call: the next selloff will be triggered not by demand but by whether capital spending converts into revenue. Two tests are close: Anthropic's IPO pricing in November, and the absolute AI revenue line in hyperscalers' next quarterly reports. Until then, every uptick in yields will be read as a repricing of AI — which is why a compute shortage and a compute bubble can coexist in the same week.

Sources: koreatimes.co.kr

Sources: kfgo.com | dw.com | startupfortune.com | cls.cn | sina.cn

The real story this week was not which model got better. It was that AI is now being asked to prove itself — that the books balance, the liability is carried, the factory floor is reachable, and the money comes back.