Policy & Regulation Story 1 of 12

Trump and Six AI Leaders Sign a Voluntary Accord, and Washington Starts Calling AI Super Intelligence

The White House has chosen its answer to the question of who polices the most powerful AI systems, and for now the answer is the companies themselves. On September 29, President Trump signed a voluntary accord with Google CEO Sundar Pichai, Anthropic CEO Dario Amodei, Meta CEO Mark Zuckerberg, OpenAI President Greg Brockman, Elon Musk and NVIDIA CEO Jensen Huang. The signatories committed to internal controls that monitor frontier models and to partnerships with independent external auditors who check that those safeguards work as intended.

The accord is nonbinding. It carries no enforcement mechanism, and the obligations rest on each company's willingness to build and fund the oversight it has promised. The structure it describes is layered: controls inside the training and deployment pipeline, a dedicated internal team responsible for those controls, outside auditors who verify them, and a standing forum where the signatories meet to agree on shared standards.

The same day, the president signed an executive order directing federal agencies to refer to artificial intelligence as "super intelligence" in official communications. The order gives the assistant to the president for science and technology 60 days to propose legislative language defining the term, which means the rebrand could eventually reach statute rather than stay on agency letterhead.

For executives, the practical signal is the absence of a regulator. The administration has declined to stand up a licensing body or a mandatory testing regime, and has instead placed the burden of proof on the labs. That shifts risk onto buyers. If model assurance is a matter of voluntary audit, enterprises deploying frontier systems will need their own evidence that controls exist and function, rather than relying on a government certificate that does not exist.

The accord also sets a template other frontier developers will be pressed to join. Any lab outside the signatory group now faces a simple question from customers and investors: why have you not made the same commitments? Expect procurement teams to begin asking vendors whether they have signed, who their external auditor is, and what the auditor is allowed to see.

The weakness is the one every voluntary regime shares. Its credibility depends on what the independent auditors are permitted to examine and whether their findings ever become public. Until those terms are written down, the accord is a statement of intent. Boards should treat it as a floor that vendors have promised to meet, and continue to ask for the documentation that proves they have.

White HouseAI SafetyFrontier ModelsSelf Regulation

AI Models Story 2 of 12

OpenAI Ships GPT-6.1 Sol at One Fifth of Astra's Price

OpenAI has released a model designed to make its flagship economics look very different. GPT-6.1 Sol, announced on September 29 at DevDay in San Francisco, costs $2 per million input tokens and $10 per million output tokens. That is one fifth of GPT-6 Astra's standard token prices, and OpenAI is positioning Sol as a near equal to Astra for the work enterprises actually pay for: coding, computer use and professional tasks.

The company's own benchmark claim is specific. On OSWorld 2.0, a test of agents operating a real computer, OpenAI says Sol comes within 2.1 percentage points of Astra at roughly one seventh of the cost per task. That second figure matters more than the headline price, because agentic workloads consume tokens in long loops and the cost per completed task is what shows up on an invoice.

OpenAI's developer documentation lists a context window of 1,050,000 tokens and up to 128,000 output tokens, which puts whole repositories and long document sets within a single request. Sol is available through the API and to ChatGPT Plus, Pro, Business, Enterprise and Edu users in ChatGPT Work and Codex.

The launch lands in an unusual moment for OpenAI's model line. The company has already set aside its planned GPT-6.1 Astra, and Sol now carries the 6.1 generation on its own. That makes Sol less a budget option than the default recommendation for most production work, with Astra reserved for the tasks where the last few points of capability justify five times the cost.

For technology leaders, the decision framework is familiar but the ratio is new. A one fifth price for near parity invites a straightforward exercise: route a representative sample of your current Astra traffic to Sol, measure task success rather than benchmark scores, and move every workload where the gap is invisible. Teams that do this well typically find that a small share of tasks genuinely needs the top model, and the rest were overprovisioned.

The competitive consequence is pressure on every rival's midrange pricing. When the market leader offers near flagship capability at a fraction of the price, competing models priced near flagship levels need a clear capability story to hold their position. Procurement teams renegotiating AI contracts this quarter now have a new reference point, and they should use it.

OpenAIGPT-6.1 SolPricingCoding

Enterprise AI Story 3 of 12

OpenAI's Dots Put Always On Agents Into Slack, Teams and the Phone

OpenAI used DevDay to make its biggest bet yet on agents that work without being asked. On September 29 the company launched dots, always on agents powered by GPT-6 Astra. Each dot runs on its own cloud computer and can connect to more than 4,000 apps, and the rollout starts with ChatGPT Pro and Business Premium users. Users can reach their dots through ChatGPT, through Slack and Microsoft Teams, and by voice call.

The design choice to give every agent its own machine is the important one. An agent with a dedicated, persistent computer can keep working after the user closes a browser tab, pick up tasks overnight, and maintain state across days. That is the difference between an assistant that answers questions and a digital worker that holds a queue.

OpenAI paired the launch with speed and pricing changes aimed at heavy users. It introduced Ultrafast, which delivers up to 8 times faster token generation in Codex, about 300 tokens per second, and up to 6 times faster generation in the API. A new Pro 500 plan costs $500 per month and includes Ultrafast access. The company also said ChatGPT has 1.2 billion weekly users, the distribution base it is now asking developers to build on.

For enterprises, dots raise governance questions faster than they answer productivity ones. An agent that acts proactively across thousands of connected applications needs clear permission boundaries, audit logs that a security team can read, and a defined owner when something goes wrong. The value case is real: recurring research, reporting, scheduling and follow up work consumes large amounts of skilled time. But the operating model for an autonomous agent inside Slack or Teams is closer to onboarding a contractor than installing software.

The practical first step is scoping. Leaders should identify a small set of repetitive, well documented workflows with low blast radius, assign each dot a named human owner, and require read only access until the agent has a track record. Connecting an agent to every app on the first day is the fastest way to create an incident.

Dots also sharpen the competitive picture. Consumer and workplace agent products from several large platforms are now converging on the same promise of persistent, proactive help. The winner will not be the agent with the most integrations. It will be the one that enterprises trust to act unsupervised, and that trust is earned in controls, not in demos.

OpenAIAI AgentsDevDayProductivity

Funding & Investment Story 4 of 12

OpenAI Is in Talks to Raise $30 Billion at About $1.4 Trillion, Bloomberg Reports

OpenAI's private market value may be about to reach a level no startup has approached. Bloomberg reported on September 29 that the company is in talks to raise at least $30 billion at a valuation of about $1.4 trillion. The round was described as a bridge to an eventual public offering rather than a replacement for one.

The scale of the step is clear from OpenAI's recent history. In March 2026 the company raised $122 billion at an $852 billion valuation. A round at roughly $1.4 trillion would mark a sharp rerating in about six months, and it would come as OpenAI's chief executive, Sam Altman, has ruled out an initial public offering in 2026.

That combination tells investors something about how the largest AI developers now finance themselves. Training and serving frontier models requires capital on a scale that used to belong only to public companies and governments. OpenAI has chosen to keep raising privately, at rising valuations, while deferring the disclosure and quarterly scrutiny that a listing brings. Its closest rival is taking the opposite path toward the public markets, which gives investors an unusual side by side comparison of the two approaches over the coming year.

For corporate buyers, the funding matters less as a headline than as a signal of staying power. Enterprises making multiyear commitments to a model provider are, in effect, underwriting that provider's ability to keep investing in compute and research. A successful raise at this level would ease concerns about OpenAI's capacity to fund its roadmap, while also raising the revenue expectations the company must eventually meet to justify the price.

There is a caution for boards as well. Valuations of this size assume that AI spending continues to grow at an extraordinary pace and that a small number of providers capture most of it. If either assumption weakens, pricing power could shift quickly toward customers. Leaders negotiating long AI contracts should build in flexibility, including model portability and exit terms, rather than assuming today's market structure is permanent.

The talks are at an early stage and terms may change before any round closes. But the direction is plain. Private capital remains willing to fund frontier AI at almost any scale, and the companies receiving it are in no hurry to trade that capital for the discipline of public markets.

OpenAIValuationPrivate MarketsIPO

Industry Dynamics Story 5 of 12

AMD Buys Fei-Fei Li's World Labs for $8.2 Billion to Build Beyond Chips

AMD is buying its way into the model business. On September 28 the chipmaker announced an agreement to acquire World Labs, the spatial intelligence startup cofounded by Fei-Fei Li, in an all stock deal valued at $8.2 billion. The transaction is expected to close by the end of 2026, subject to regulatory approvals, and Li will become AMD's Executive Vice President and Chief Scientist.

World Labs builds models that generate, reconstruct and simulate interactive three dimensional environments from text, images and video, along with technology aimed at robotic learning and simulation. That is a different frontier from the language models that dominate enterprise AI today. Spatial models are the foundation for robots that need to understand the physical world, for digital twins of factories and warehouses, and for simulation environments where autonomous systems can be trained safely before they touch real equipment.

The strategic logic for AMD is about the full stack. Its largest rival has spent years building software, models and developer ecosystems around its hardware, which makes its chips harder to replace. By bringing a leading research team in house, AMD gains the ability to shape the workloads its accelerators will run, rather than simply competing on price and performance for workloads designed elsewhere. Paying in stock also preserves cash for the capacity buildout the AI market continues to demand.

For enterprises, the deal is a reminder that the AI infrastructure market is consolidating around vertically integrated platforms. When a chip company owns the models, the tooling and the research agenda, customers get tighter optimization but also deeper lock in. Technology leaders planning physical AI programs in robotics, logistics or industrial simulation should expect hardware vendors to increasingly bundle models and software with silicon, and should evaluate those bundles on portability as well as performance.

The acquisition also signals where the next competitive contest is heading. Language models have become widely available and increasingly priced as a commodity. Models that understand space, motion and physical cause and effect remain scarce, and the teams that build them are few. AMD has decided that owning one of those teams is worth more than $8 billion.

The open question is retention and focus. Research labs acquired by large companies can lose momentum as priorities shift. How AMD structures World Labs inside its organization, and whether its research stays open to the broader ecosystem, will determine whether this becomes a platform advantage or an expensive talent acquisition.

AMDWorld LabsAcquisitionsSpatial Intelligence

Policy & Regulation Story 6 of 12

Appeals Court Rules Training AI on Westlaw Headnotes Was Not Fair Use

The first federal appeals court to rule on fair use in AI training has sided with the copyright holder. On September 29 the U.S. Court of Appeals for the Third Circuit affirmed that ROSS Intelligence's use of Thomson Reuters' Westlaw headnotes to build an AI legal research tool was not fair use. The opinion, written by Judge Tamika Montgomery-Reeves, remained under seal pending redactions, leaving the precise reasoning unpublished even as the outcome became clear.

The case began in the district court, where Judge Stephanos Bibas, sitting by designation, found that ROSS infringed 2,243 Westlaw headnotes. ROSS had used the material to train a system that competed directly with Westlaw, a fact that weighed heavily against the startup because the use served the same market as the original work.

For companies building or buying AI, the ruling moves the copyright question from theory into precedent. Dozens of cases over generative AI training are working through the courts, and until now no appellate court had addressed whether copying protected material to train a model qualifies as fair use. This decision does not settle those cases, because ROSS was a narrow dispute about a legal research tool built to compete with the very product it copied. But it establishes that training use is not automatically protected, and that the competitive relationship between the AI product and the source material matters.

The practical exposure is concentrated in vertical AI. Legal, financial, medical and data publishing businesses have spent decades building proprietary editorial content. AI vendors that trained on that content to build competing products now face a clearer legal risk than they did a week ago. Enterprises buying specialized AI tools should ask vendors directly what data trained their systems, whether it was licensed, and whether the vendor will indemnify customers against infringement claims.

The ruling also strengthens the hand of content owners in licensing negotiations. Publishers with valuable, structured content now have an appellate decision to point to when setting terms, which is likely to push more AI developers toward paid data deals rather than scraping.

Much depends on what the sealed opinion says when it is released. Whether the panel adopted the district court's full reasoning, or decided on narrower grounds, will shape how far this precedent reaches. General counsel should track the unsealing and review their AI data sourcing policies in the meantime.

CopyrightFair UseThomson ReutersLegal AI

AI Safety Story 7 of 12

Anthropic Says an Open Weight Chinese Model Now Builds Exploits Nearly as Well as Its Own

Advanced offensive cyber capability is no longer confined to tightly controlled frontier models. In research published September 29, Anthropic reported that Zhipu's open weight GLM-5.3 model developed end to end exploits in 50 of 410 attempts on ExploitBench, compared with 56 of 410 for Anthropic's own Claude Mythos Preview. Anthropic keeps its model behind safeguards and restricted access. GLM-5.3 can be downloaded by anyone.

The safeguard findings are the more troubling half of the report. Anthropic found that GLM-5.3 engaged with harmful cyberattack requests 64% of the time when given a false cover story, 92% of the time with prefilled reasoning, and 100% of the time in an abliterated version, a modified copy with its refusal behavior stripped out. Anthropic estimated that removing the model's safeguards through abliteration cost about $4,400 in compute, roughly 2,200 GPU hours.

The cost of attack is falling as well. In one test, a researcher used GLM-5.3-Flash, a smaller version of the model, to build a working exploit for Chrome vulnerability CVE-2026-11645, a flaw that had already been publicly disclosed and fixed. It took 20 minutes of human attention and 8 hours of model time, at a cost of $20.40 at Zhipu's API prices.

For security leaders, the implication is direct. The window between a patch being published and an attacker weaponizing it is shrinking, and the tools that shrink it are now freely available. Patch management cycles measured in weeks were built for an era when exploit development required scarce human expertise. That assumption no longer holds, and critical systems facing the internet need patching measured in days.

Anthropic's own recommendation is that defenders should use the best available tools that meet their needs, and that governments should conduct safety testing on sufficiently capable models. The first half is actionable today. Security teams that are not yet using AI to find and fix vulnerabilities in their own code are ceding ground to attackers who already can.

The report also adds weight to the policy debate over open weight releases. Controls on a single company's model mean little when a comparable capability is published openly. Whatever governments decide, enterprises cannot wait for that debate to conclude. The prudent assumption is that capable, unrestricted offensive AI is already in the hands of adversaries, and defense budgets should be planned accordingly.

AnthropicGLM-5.3CybersecurityOpen Weight Models

AI Safety Story 8 of 12

Microsoft Details an Agent Driven Attack That Deleted Azure Storage in Minutes

Microsoft has published a detailed account of what automated, agent driven destruction looks like inside a cloud tenant. Microsoft Threat Intelligence reported on September 25 that Storm-3168, an actor linked to the JADEPUFFER agentic ransomware operation, used compromised Azure service principals to attack a customer environment. The destructive phase was fast: a sequence of more than 100 storage account deletion attempts in about seven minutes.

The entry point was ordinary. Microsoft said the service principal credentials the attackers used had previously been exposed in plaintext in a public GitHub issue. No novel exploit was required. A leaked secret, with enough permissions attached, was sufficient to give an automated attacker the keys to production infrastructure.

What makes the case significant is the speed of the final stage. A human operator deleting resources one by one gives defenders time to notice and respond. An automated sequence that moves through storage accounts, databases and key vaults in minutes leaves almost no window. By the time an alert reaches an on call engineer, the damage can already be done.

For technology leaders, the lessons are about identity rather than artificial intelligence. Service principals and other machine identities are multiplying across every cloud estate, often with broad permissions granted for convenience and rarely reviewed. Each one is a credential that, if leaked, lets an attacker act with the full authority of an automated system. Organizations should inventory these identities, apply least privilege, rotate secrets regularly, and scan public code repositories for exposed credentials.

Recovery design matters just as much. Resource locks, immutable backups held outside the primary tenant, and deletion protections turn a catastrophic incident into a recoverable one. Microsoft's guidance emphasizes safeguarding backup and recovery infrastructure precisely because destructive attackers target it first.

The broader point is that agentic tools cut both ways. The same automation that lets companies run operations at machine speed lets attackers destroy them at machine speed. Defense has to match that tempo, which means automated detection and response that can halt a destructive sequence without waiting for human approval. Security teams still relying on manual triage for cloud control plane alerts should treat this report as a timeline test: could your team stop 100 deletions in seven minutes?

MicrosoftCloud SecurityAgentic AttacksIdentity

Generative AI Story 9 of 12

Washington Launches America.gov, an AI Front Door to Federal Services

The federal government now has a single conversational entry point. On September 29 the Trump administration launched America.gov, an AI chatbot for federal services powered by Google Gemini and xAI Grok. The project is led by Joe Gebbia, the U.S. Chief Design Officer and a cofounder of Airbnb. FedScoop reported that the site draws on information from approximately 29,000 government websites.

The idea is simple and overdue. Citizens looking for a passport, a campsite reservation or a replacement benefits card have long had to guess which agency and which website held the answer. A single interface that understands plain language questions and routes people to the right service removes much of that friction. At launch the system points users to the relevant services rather than completing transactions itself.

The choice of two commercial models from two different companies is notable. It suggests the government is treating foundation models as interchangeable components rather than committing to a single vendor, an approach many enterprises are also adopting to avoid lock in and to compare performance on real traffic.

For enterprise leaders, America.gov is a large scale case study in customer service consolidation. Many companies face the same problem the government does: thousands of pages of product, policy and support content spread across disconnected sites, with customers who cannot find what they need. The pattern here, an AI layer that reads existing content and routes questions without replacing the underlying systems, is one of the fastest ways to improve customer experience without a multiyear platform rebuild.

The risks are also instructive. A government chatbot answering questions about benefits, taxes or legal status carries real consequences when it is wrong. Accuracy on authoritative source content, clear escalation to human help, and transparency about data handling will determine whether the public trusts it. Those same requirements apply to any company putting AI in front of customers on regulated topics.

Privacy and procurement questions will follow. Citizens will want to know what happens to their questions and how the commercial providers are contracted. Businesses watching the rollout should note that public scrutiny of this kind of deployment is intense, and that the governance choices made here are likely to shape expectations for private sector AI service as well.

Government AIAmerica.govGeminiGrok

AI Business Models Story 10 of 12

EliseAI Raises $350 Million at a $4 Billion Valuation for Housing and Healthcare AI

Some of the strongest AI businesses are being built far from the frontier labs. EliseAI, which automates operations in housing and healthcare, announced on September 29 that it raised $350 million at a $4 billion valuation, in a round led by Andreessen Horowitz and Bessemer Venture Partners.

The company's product handles the unglamorous, high volume work that consumes staff time in property management and medical practices: leasing inquiries, maintenance requests, patient scheduling and billing. The company says it had more than $200 million in annual recurring revenue as of June 2026 and that its platform powers one in six U.S. apartments. Those figures describe a business built on repeatable workflows rather than on model breakthroughs.

That is the lesson for executives weighing where AI value actually accrues. Much of the public conversation focuses on the companies training ever larger models. EliseAI's growth reflects a different strategy: pick industries with enormous volumes of routine communication, deep regulatory and operational complexity, and chronic staffing shortages, then build agents that fit precisely into existing workflows. The model underneath can change. The workflow knowledge, integrations and customer relationships are the durable asset.

For operators in housing and healthcare, the round signals that AI automation of front office work is moving from pilot to standard practice. Competitors that still route every tenant question or appointment change through human staff will face a growing cost disadvantage. The question is no longer whether to automate these interactions, but how to do it without degrading service quality or creating compliance exposure in sensitive areas such as fair housing and patient privacy.

For investors and corporate strategy teams, vertical AI valuations are now high enough to reward careful diligence. Recurring revenue at this scale is meaningful, but so is concentration risk if a platform becomes dependent on a small number of large property managers or health systems. Retention, expansion within existing customers, and the ability to add new workflows will determine whether the valuation holds.

The broader takeaway is that the AI economy is widening. The most transformative near term deployments may come from companies that understand one industry deeply and apply capable models to its daily grind. Enterprises should look for those specialists in their own sectors before trying to build everything in house.

EliseAIVertical AIFundingHealthcare

AI Infrastructure Story 11 of 12

Efficient Computer Raises $97 Million to Take Its Low Power Chips From the Edge to the Data Center

Energy, not raw speed, is becoming the defining constraint in AI hardware, and investors are funding companies that attack it directly. Efficient Computer said on September 29 that it raised a $97 million Series B at a $650 million valuation, led by TQ Ventures, bringing its total funding to $173 million. The company's Electron E1 processor is in volume production.

Efficient Computer builds processors around a dataflow architecture designed to cut the energy consumed by general purpose computation. Its first market is physical AI: robots and other devices where battery life and heat, not peak performance, determine what a product can do. The new funding is intended to scale production and to extend the architecture toward data center class performance.

The timing reflects a broader shift in the infrastructure market. Data center expansion is increasingly limited by power availability, and at the edge, devices face hard limits on how much energy they can draw. A chip that performs the same work on a fraction of the power opens applications that were previously impractical, from longer running autonomous robots to always on sensing in remote locations.

For technology leaders, the relevance is in planning horizons. Many physical AI projects stall not because the models fail but because the compute required to run them does not fit the power budget of the device. Specialized, energy efficient processors are beginning to close that gap. Teams building robotics, industrial monitoring or connected equipment should revisit designs that were ruled out on power grounds, because the hardware options are changing quickly.

The data center ambition is the harder test. Moving from edge devices into the data center means competing against entrenched accelerator ecosystems with mature software, tooling and developer communities. New architectures often win on efficiency benchmarks and then struggle with adoption because customers cannot easily port their existing workloads. Efficient Computer's success there will depend as much on its compiler and software stack as on its silicon.

Still, the round is a signal worth noting. As AI spreads from chat interfaces into physical machines, the economics of every watt start to matter. Companies that can deliver useful computation at dramatically lower energy cost will find customers well beyond the headline AI labs, and the enterprises that plan around them early will have more options than those that do not.

Efficient ComputerSemiconductorsEnergyPhysical AI

AI Infrastructure Story 12 of 12

Accelevation Prices Its IPO Below Range, a Rare Cool Signal for AI Infrastructure

Public investors are showing they will still negotiate on price, even for companies riding the data center boom. Accelevation, which makes power distribution and thermal equipment for data centers and is controlled by private equity firm Olympus Partners, sold 30 million shares in its Nasdaq initial public offering after marketing them at $20 to $24 each. IPO Scoop reported that the deal priced at $18 per share, below the range, raising $540 million for the company and selling shareholders.

The business sits squarely in the path of AI spending. Every new AI cluster needs power delivered to racks and heat removed from them, and Accelevation supplies the equipment that does both. Its registration filing reported a backlog of about $1.1 billion, a sign of the demand flowing through the physical layer of the AI buildout.

Pricing below the range despite that backlog is the notable part. It suggests investors are distinguishing between the AI boom as a theme and the valuation they will pay for any individual company exposed to it. A private equity owner selling a large block of shares alongside the company, and a business tied to a spending cycle that could slow, give buyers reasons to demand a discount.

For executives, the signal is about discipline rather than retreat. Capital is still flowing to AI infrastructure, but public markets are beginning to price in cyclicality. That matters for companies planning their own infrastructure investments and for suppliers weighing expansion. The assumption that every AI adjacent asset commands a premium is being tested, and lenders and investors are likely to scrutinize the durability of demand more closely in coming quarters.

It also offers a useful lens on supply chains. Power and cooling equipment has been one of the tightest bottlenecks in data center construction. A healthy backlog at a key supplier indicates that constraint has not eased. Enterprises planning on premises AI capacity, or depending on colocation providers to deliver it, should still expect long lead times for electrical and thermal gear and should lock in supply early.

The first weeks of trading will show whether the discount was prudence or pessimism. Either way, the offering is a reminder that the physical infrastructure behind AI is a real industry with real cycles, and that the market is starting to treat it that way.

AccelevationIPOData CentersPower