Past Issues
The hardest problem in AI right now is management, not capability. Google shipped another mid-tier Gemini while its flagship stays late and its most celebrated researchers leave. OpenAI is losing senior executives while signing a data centre financing arrangement worth up to $105 billion. Anthropic raised its own catastrophic misalignment rating from very low to low three days before its CFO started taking early IPO meetings. And xAI now sits inside SpaceX. The models keep getting better. The companies building them are getting harder to run.
THE MACRO
Google’s head start never became a model business.
Google helped invent the architecture underlying the generative AI boom. The transformer originated there. DeepMind became one of the world’s most important research laboratories. Google had proprietary datasets, unlimited distribution, one of the largest cloud businesses, extraordinary compute and some of the best AI researchers alive. It also released enormous amounts of foundational research openly, which helped create the ecosystem that now competes with it.
Google remains one of the most important AI companies in the world, and its strongest commercial position is enterprise infrastructure rather than the model layer. Vertex AI lets businesses access and orchestrate multiple model families instead of forcing customers into Google’s models exclusively. That is a powerful business. It is also a different victory from owning the models themselves.
Open weights are testing the price of intelligence.
Chinese open-weight models are increasingly capable, increasingly cheap, and can be downloaded, modified and run independently. American enterprises are beginning to use them. Chinese models now occupy the top five positions by weekly token usage on OpenRouter.
The best closed models retain real capability advantages. Most enterprise AI tasks don’t need them. They need a model that is good enough, reliable enough and cheap enough. OpenAI, Anthropic and Google are spending extraordinary sums to maintain capability leadership while their competitors give capable intelligence away. The operative question has shifted from who builds the smartest model to how much customers will keep paying for the smartest model once sufficiently capable alternatives are cheap or free.
The pricing already shows this. Gemini 3.7 Flash launched this month at half the introductory rate of the model it replaced, three weeks after that model shipped. DeepSeek’s V4 Flash is listed at roughly a fifth of Google’s introductory price. Nobody cuts prices that fast in a market they control.
Compute is being financed on somebody else’s credit.
Nvidia will provide up to $105 billion in financing for OpenAI’s Ohio data centre campus, disclosed in a securities filing on August 17. The credit supports an initial 4.25 gigawatts with an option for another 3.75, SB Energy builds and manages the site, and OpenAI signs a 20-year lease with capacity arriving in phases from 2028.
The structure matters more than the number. OpenAI does not have an investment-grade credit rating, so the debt gets raised against Nvidia’s balance sheet instead. The chip vendor is now underwriting its own customer’s ability to buy chips. Earlier reporting put the guarantee as high as $250 billion before it was cut back, which tells you the number is a negotiation rather than a plan.
The most defensible AI company may not be an AI company.
SpaceX’s absorption of xAI gives investors a publicly traded platform that contains a lab building frontier models. Grok has not established anything resembling ChatGPT’s consumer position or Anthropic’s enterprise position. Its advantage is that it is embedded: it lives inside X, inside a company with enormous distribution, infrastructure, connectivity and capital. If intelligence itself commoditizes, the durable position belongs to the platform that contains the intelligence rather than the lab that produces it. Google appears to be moving toward that conclusion. SpaceX arrived there by acquisition. OpenAI and Anthropic still have to prove the model company can remain the destination.
THE MICRO
Google shipped Gemini 3.7 Flash on August 13, twenty-three days after 3.6 Flash. DeepSWE v1.1 moves from 49.0% to 65.3% on Google’s own evaluations, FrontierCode 1.1 Main from 34.4% to 43.6%, and WebDev Arena Elo from 1538 to 1588. Context window stays at roughly one million tokens. Introductory pricing is $0.75 per million input tokens and $3.75 output through the end of 2026, half what 3.6 Flash charged at launch, rising to $1.50 and $7.50 on January 1, 2027. Google attributes the gains to algorithmic work and developer feedback rather than a larger model.
Gemini 3.5 Pro, promised for June, still has no release date. Google also began retiring three Imagen 4 model IDs on August 17 and is directing developers to gemini-3.1-flash-image, where the old generate_images() method is gone and image generation runs through generate_content(). Anyone with Imagen 4 endpoints in production needs to re-test prompt adherence, aspect ratios, SynthID handling and quotas.
Google’s people. Demis Hassabis is stepping away from day-to-day management of DeepMind to become Alphabet’s chief scientist and DeepMind chairman, with Koray Kavukcuoglu taking greater operational authority. Jeff Dean, Sanjay Ghemawat, Quoc Le and Oriol Vinyals have left to form Discovery Loop. Noam Shazeer, one of the authors of the transformer paper, left for OpenAI. Nobel laureate John Jumper left for Anthropic.
Anthropic published its second company-wide risk report on August 14, a 186-page document under version 3.4 of its Responsible Scaling Policy covering February 24 through a July 15 coverage date. The rating for catastrophic harm from misalignment in high-stakes settings moved from very low to low. Anthropic attributes the change to increased overall uncertainty following recent cybersecurity evaluation disclosures, and says its underlying argument probably still supports the lower rating.
The report also discloses an internal model, called Model 2, that Anthropic describes as somewhat more capable than Mythos 5 and heavily used inside the company, with no current plans for external release and no full predeployment assessment suite run against it. Risk from non-novel biological and chemical weapons uplift stays low, described as higher than the previous estimate, after the company found that all human-feedback vendor traffic between May 2025 and April 2026, roughly 133 million exchanges with about 50,000 contractors, ran without its biological-weapons blocking classifiers active. Anthropic says the gap is remediated, no customers were affected, and its review found no evidence of harmful misuse, while noting the discovery reduced its confidence that no similar gaps exist. Its internal benchmark for the automated AI R&D threshold has saturated and can no longer register incremental capability gains.
OpenAI’s people and money. Brad Lightcap, one of Sam Altman’s longest-serving lieutenants, departed this month. Chief Revenue Officer Denise Dresser has also left. Fidji Simo has stepped away from her full-time leadership role for health reasons. Those departures follow a longer list across product, safety, research and enterprise. OpenAI has also disbanded its Preparedness team and redistributed its responsibilities. Separately, the company is funding 14 policy and research projects examining AI’s effects on employment, economic opportunity and public policy, with $1 million in cash plus up to another $1 million in model credits split across recipients spanning the US political spectrum and organizations in Europe, Brazil, Singapore and South Korea.
Everyone else shipped too. Alibaba released Qwen3.8 Max on August 2 and Qwen3.8-27B on August 14. DeepSeek moved DeepSeek-V4-Pro-0813 to general availability across app, web and API on August 13, with V4-Flash-0731 still in public beta and no sign of V5 or R2. Z.ai released GLM-5.3 on August 14. ByteDance released Seed 2.1 Turbo on August 10. xAI released Grok Imagine Image 2.0 on August 8, and Grok 4.6 posted another significant benchmark jump. Databricks closed a $5 billion round at a $190 billion valuation.
Got your tickets for MAICON 2026, October 13 to 15 in Cleveland yet?
The Marketing AI Conference returns to the Huntington Convention Center of Cleveland for its seventh year, hosted by the Marketing AI Institute. Three days on what marketing teams are actually deploying: tool evaluation and ROI measurement, content operations, personalization, governance and change management, and case studies from CMOs and operators running AI inside their organizations.
The AI for CMOs Summit runs the morning of October 13, with half-day workshops that afternoon. Main conference keynotes and breakouts are October 14 and 15. Passes come as Main Conference, All-Access with the workshop, or All-Access plus on-demand main stage recordings. Groups of four or more get 10% off, and there are custom rates above ten.
B2BNN is a media partner of MAICON 2026. https://hubs.li/Q049vD6T0 Use code AKRON150 for a discount on your registration!
THE POLICY
The EU’s transparency obligations are live and the technology is behind them. Article 50 of the AI Act became enforceable on August 2, covering chatbot disclosure, deepfake labelling, and disclosure of AI-generated text published on matters of public interest unless a human holds editorial responsibility. The Commission adopted its final 51-page Article 50 guidelines on July 20 and assessed the voluntary Code of Practice on Transparency of AI-Generated Content as adequate.
The Digital Omnibus on AI, Regulation (EU) 2026/1744, entered into force on July 27 and left Article 50 alone. High-risk obligations moved: standalone Annex III systems now bind on December 2, 2027, and AI embedded in regulated products under Annex I on August 2, 2028. Generative systems already on the EU market before August 2 have until December 2, 2026 to meet the machine-readable marking requirement in Article 50(2). Systems placed on the market after August 2 comply now.
The Commission’s own guidelines acknowledge that technical standards for measuring compliance with the marking obligation are still being developed through the Code of Practice and EU standardization work. The law requires machine-readable marking and is technology-neutral about method. Watermarks are the practical option for audio, image and video, and they degrade under compression, cropping, re-encoding and paraphrase. How individual market surveillance authorities read Article 50 will decide what compliance means in practice.
California is doing the work Washington isn’t. Roughly 30 AI bills went through appropriations suspense votes in both chambers on August 13, including measures for a state-specific auditing and standards system, call centre oversight, and restrictions on AI-assisted employment decisions. Twenty-seven states have passed 85 new AI-related laws so far in 2026, and seven legislatures remain in session: California, Michigan, Pennsylvania, Massachusetts, Ohio, New Jersey and North Carolina.
Federal preemption is still mostly on paper. Executive Order 14365 directed the DOJ to stand up an AI Litigation Task Force, the FTC to issue a policy statement on state laws requiring alterations to model outputs, and Commerce to publish a review of onerous state AI laws by March 11, conditioning leftover BEAD broadband funds on repeal. The task force was announced in January. The Commerce evaluation is more than five months overdue. Industry keeps asking for the federal framework anyway: the American Bankers Association’s response to House Financial Services Committee Democrats this week calls for a federal AI framework that preempts state regulation.
Frontier model pre-release review is becoming a habit. Reporting on the White House’s ad hoc review process, formalized after the June export controls on Anthropic’s Fable 5 and Mythos 5, describes the Office of the National Cyber Director meeting with Anthropic, Google, Meta and OpenAI to explain how the framework works. Open-weight models are excluded. Companies are encouraged to share covered proprietary models with the government as close to public release as possible. An informal process is turning into a structure without a statute behind it.
THE GOSSIP
The doom slayers are extending their horizons. The AI bubble burst call keeps sliding. On Polymarket, the “AI bubble burst by” market had traded around $2.9 million by August 9, with December 31, 2026 as the most-backed single date at roughly 12%, and the headline probability drifting from about 26% in June to 18 or 19% at the end of July. The analyst consensus has softened from a collapse to a 20 to 30% correction in AI-heavy equities spread across 2026 and 2027. Ed Zitron now allows that a burst would be a succession of events taking upwards of a year, and one VC’s arithmetic on burn rates puts full collapse around February 2027.
US AI equities are priced for perfection, hyperscaler capex guidance is around $725 billion for 2026 against $410 billion last year, and every quarter this year raised the number rather than trimming it. A repricing of American AI stocks is a plausible event. It is a different event from AI failing to work, and a different event again from the technology’s centre of gravity moving. Chinese labs are shipping capable models at a fifth of American prices and taking the top five OpenRouter slots by usage while the burst forecast slides another year to the right. A US market failure and an AI failure are not the same story.
The $2 trillion number is coming from somewhere other than Anthropic. Reports put a prospective Anthropic IPO at roughly $2 trillion, which would be the largest offering in history. CFO Krishna Rao is leading early meetings with investors and reportedly has not discussed valuation in them. There is also a steady stream of criticism circulating about Dario Amodei’s leadership and Anthropic’s safety positioning, much of it less substantive than the documented personnel changes at Google and OpenAI. The timing is the interesting part regardless: the company raised its own misalignment risk rating and disclosed a classifier gap covering 133 million exchanges in the same month its CFO started taking investor meetings. That is either unusual discipline or unusual exposure, and the S-1 will settle which.
Analysts are reading the Flash cadence as a talent story. Two Flash releases in three and a half weeks while Gemini 3.5 Pro stays two months past its promised date has people arguing that Google’s mid-tier is where the engineering now lands because the people who shipped the top tier have left. Google hasn’t said anything about the delay.
The off-balance-sheet question. The reporting this week that Big Tech may have trillions more committed to AI infrastructure than balance sheets show is the version of the bubble argument worth watching, because it is checkable. Lease obligations, vendor guarantees and take-or-pay compute contracts are where a capex slowdown would first become visible.
WHAT TO WATCH
- Xi in Washington on September 24. The UN General Assembly’s General Debate opens September 22 in New York, and reporting this week says Xi is expected to skip it, arriving in the US on September 23 for a one-day White House meeting with Trump on the 24th and leaving on the 25th. It would be his first US trip in 11 years and his first in-person appearance at UNGA since 2015 will not happen. Trade and Taiwan are the named agenda items. Anything said in that room about chips and export controls will move more for AI than the entire September conference calendar, and skipping the General Assembly means China makes its case bilaterally rather than to the 193.
- AI for Defense Summit, September 2 and 3 in Washington. Fifth annual, run by DSI Group at the AIA Global Campus, convening Department of War, intelligence community and industry on moving AI from experimentation to deployment. The procurement language that comes out of it is the tell for allied buyers.
- Government & AI Summit, September 15, JW Marriott on Pennsylvania Avenue, on federal deployment, AI-powered cybersecurity and governance.
- Unitree lists on Shanghai’s STAR Market on August 19, the first general-purpose robotics company to go public on mainland China’s markets.
- Gemini 3.5 Pro. Whether it ships, and what Google says about why it took this long. Google rebuilt the base model from scratch rather than ship what it had, missed a July 17 target, and has now put out two Flash releases in the time the flagship has been late.
- The other three flagships in the queue. OpenAI named its next major model Astra on August 1 with no date and no pricing. Grok 5 is still training on Colossus 2 while xAI ships 4.x point releases and folds in Cursor, whose $60 billion acquisition by SpaceX is expected to close this quarter, with the next model built to train on Cursor data from the start. DeepSeek has published nothing about V5 or R2 and is iterating V4 instead. Four labs are all late or quiet on the top tier at the same time, and all four are shipping cheap mid-tier models at a sprint pace. That pattern is the story to write when one of them finally lands.
- Anthropic S-1 signals. Whether the risk report’s language about misalignment ratings and classifier gaps survives into investor materials.
- December 2 in the EU. Machine-readable marking binds for systems already on the market, and the first Article 50 enforcement action from any market surveillance authority will set the tone.
- California floor votes on the bills that cleared suspense, before the session ends.
- Commerce’s overdue state-law review, and whether the AI Litigation Task Force files anything.
- Whether another compute deal gets financed on a vendor’s credit rating. The Nvidia and OpenAI structure is now a template.
