Today's 12 Stories — Friday, September 25, 2026

Policy & Regulation Story 1 of 12

Altman and Amodei Ask the UN Security Council for Global AI Rules, and Washington Says No

The chief executives of the companies building the most capable AI systems went to the most powerful room in international diplomacy this week and asked to be governed. The United States government, sitting at the same table, told them it wants no part of it.

The UN Security Council held a briefing on artificial intelligence on September 23. Speakers included OpenAI CEO Sam Altman, Anthropic CEO Dario Amodei, Hugging Face CEO Clement Delangue and the AI researcher Yoshua Bengio. The session took place in New York during the General Assembly's high level week. It was an unusual lineup for a body that normally spends its time on wars, sanctions and peacekeeping mandates, and the choice of venue was itself the message: the people closest to frontier AI now describe it as a matter of international security rather than a technology policy question.

Amodei made the most concrete ask. He called for common global standards for testing AI models for loss of control risks and misuse risks, along with systems for reporting incidents when things go wrong. That is a recognizable regulatory architecture, borrowed from aviation and nuclear safety, in which the testing regime and the incident log matter more than any single rule.

Altman framed the issue around legitimacy. He told the Council that AI rules must be shaped through democratic processes and by governments accountable to the people they serve. Read carefully, that is both an endorsement of regulation and a quiet argument about who should write it, with national governments rather than industry or unelected international staff holding the pen.

The Trump administration answered bluntly. White House science adviser Michael Kratsios said the United States rejects efforts by international bodies to assert centralized control over AI. That position is consistent with the administration's broader approach, which treats AI leadership as a strategic contest to be won through speed and domestic investment rather than a shared risk to be managed through treaty.

For executives, the split matters more than the speeches. The companies at the frontier are now publicly on record asking for binding testing and incident reporting standards, while their home government refuses to negotiate them multilaterally. That gap will not stay empty. Expect the European Union, the United Kingdom and coalitions of middle powers to move toward testing and disclosure regimes that American companies must satisfy anyway to sell in those markets.

The practical takeaway is to plan for a patchwork. Companies deploying frontier models should assume that model evaluation evidence, incident logs and misuse controls will become procurement requirements in at least some jurisdictions, whatever happens in Washington. Building that documentation now is cheaper than retrofitting it under deadline later, and the Security Council session made clear that the vendors themselves expect it to arrive.

Policy & Regulation Story 2 of 12

Sanders and Casar Introduce a Bill to Ban Artificial Superintelligence Outright

The most aggressive AI bill yet introduced in the United States Congress arrived on September 23, and it does not propose to regulate advanced AI so much as to stop it.

Senator Bernie Sanders and Representative Greg Casar introduced the Ban Artificial Superintelligence Act. The bill would create a cabinet level Department of Artificial Intelligence and pause advanced AI development until that department sets safety rules. The ban itself would permanently prohibit the development of artificial superintelligence, meaning systems that would exceed human capability across most domains.

The enforcement provisions are designed to be impossible to ignore. The bill provides for dissolving companies that violate the ban and prison terms of up to 20 years for individuals. The sponsors frame those penalties as comparable to the treatment of unlawful nuclear weapons development, which is the analogy the bill's entire logic rests on: a technology whose worst case is so severe that the only acceptable policy is prohibition backed by criminal law.

The bill has almost no chance of passing the current Congress, and its sponsors surely know it. Its significance lies elsewhere. It moves the Overton window on AI policy in a direction that, until recently, lived mostly in research papers and online forums. A cabinet department dedicated to AI, a development pause tied to regulatory approval, and a corporate death penalty are now written into legislative text with named congressional sponsors, which makes them available as reference points in every future negotiation.

It also lands in the same week that AI company leaders told the UN Security Council they want global testing standards, and the administration rejected international control. The political spectrum on AI is widening fast. At one end sits an executive branch focused on acceleration and national competitiveness. At the other sits a progressive faction arguing that the frontier labs are building something that should not be built at all. The middle ground of licensing, audits and liability, where most practical policy will likely land, is being defined by these two poles.

For boards and executive teams, the lesson is not that superintelligence will be banned. It is that AI governance has become a live electoral issue with real constituencies, and positions that once looked fringe can harden into party platform commitments within a cycle or two. Companies whose strategies depend on unconstrained scaling of frontier models should treat regulatory risk as a scenario to plan for, not a tail risk to dismiss, and should document their safety practices well before anyone in Washington asks to see them.

Policy & Regulation Story 3 of 12

UK Regulator Wants ChatGPT and Perplexity on Google's Search Choice Screens

Britain's competition regulator has decided that the future of search includes chatbots, and it wants Google's own devices and browser to say so.

The UK Competition and Markets Authority on September 23 proposed that AI assistants such as ChatGPT and Perplexity could be eligible to appear on Google's search choice screens on Android and Chrome, provided they meet technical and security criteria. The proposal strengthens an earlier plan by widening the field of eligible rivals beyond traditional search engines.

The mechanics matter. Under the revised proposal, users would see a search choice prompt when they first set up an Android device or open Chrome, and the prompt would reappear once a year. The screen would list up to 12 search providers: the five most popular plus seven chosen at random. That structure is meant to give newer entrants a real chance at visibility without simply entrenching whoever is already largest. The CMA is accepting consultation responses until October 9.

The regulatory logic is straightforward. The CMA's view is that Google's control of defaults on the world's most widely used mobile operating system and browser shapes which search services people ever try. Including AI assistants acknowledges what usage trends already suggest: for a growing share of queries, people now go to a conversational assistant rather than a list of links. A choice screen that ignored that shift would be regulating the market of five years ago.

For Google, the risk is not that a choice screen instantly moves market share. The larger risk is symbolic and cumulative. A regulator formally treating AI assistants as substitutes for search strengthens the argument that Google's dominance in traditional search should not automatically extend into AI answers, and it gives rivals a regulated distribution channel onto the devices where most people first choose how they search.

For OpenAI, Perplexity and other assistant makers, a place on the screen would be valuable distribution that money cannot easily buy. It would also bring obligations. Meeting technical and security criteria set by a regulator means accepting scrutiny over how answers are produced and attributed, which is a new kind of oversight for companies used to competing on product alone.

Enterprise leaders should watch this as a signal of where competition law is heading. Regulators are beginning to define AI assistants as a market in their own right, adjacent to and overlapping with search. That definition will shape how default placements, bundling and data access disputes are judged across the AI industry, well beyond Google.

AI Safety Story 4 of 12

NHTSA Opens Investigation Into comma.ai Driver Assistance Devices After Fatal Crashes

The federal auto safety regulator has turned its attention to a corner of the driving automation market that has largely escaped scrutiny: aftermarket AI systems that owners install in cars themselves.

The National Highway Traffic Safety Administration opened preliminary evaluation PE26007 into comma.ai's openpilot driver assistance devices on September 21, with coverage of the probe spreading this week. The investigation covers about 30,000 comma three, comma 3X and comma four devices. NHTSA identified five crashes involving the devices in which vehicles struck stopped or slow moving vehicles in their lane, including two fatal crashes that killed three people.

comma.ai occupies a distinctive position. Its devices are aftermarket units that owners install in supported cars to control steering and speed, running openpilot, which is open source driver assistance software. That model raises a hard regulatory question: when an aftermarket AI takes control of a car built by someone else, who is responsible when it fails?

The crash pattern NHTSA describes is a familiar one in driver assistance. Striking stopped or slow vehicles in the travel lane is exactly the failure mode that has drawn scrutiny to other systems, because camera based perception can struggle to distinguish a stationary obstacle from the roadway and because drivers who trust the system may react too late.

A preliminary evaluation is the first step in NHTSA's defect process. It can close without action, escalate to an engineering analysis, or lead to a recall request. Even at this stage, it changes the landscape. Aftermarket automation vendors can no longer assume they sit outside the oversight that applies to automakers.

For executives beyond the automotive sector, the case illustrates a broader principle. AI that acts in the physical world is judged by outcomes, not by architecture or intent. Open source code, a passionate user base and strong performance in normal conditions do not answer the questions a regulator will ask after a fatal failure. Companies deploying AI agents that control physical systems, from warehouses to vehicles to industrial equipment, should expect incident data to become the currency of accountability, and should build logging, failure analysis and escalation procedures before an investigator requests them.

The outcome of this probe will also help define how the United States treats retrofit AI more generally, a category that will grow as capable models become cheap enough to bolt onto existing machines.

AI Models Story 5 of 12

Google Launches Gemini 3.8 Flash TTS and Flash-Lite TTS, Its Most Expressive Voice Models Yet

Google has pushed further into synthetic voice, releasing two new text to speech models this week that aim squarely at developers building voice agents, audio content and localized media.

Google announced Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS, describing them as its most expressive audio generation models to date. Google said the models give access to more than 2,000 production ready voices, and that Gemini 3.8 Flash TTS can create voices in more than 100 languages and dialects. The larger Flash model targets quality and expressiveness, while Flash-Lite is positioned for cost sensitive, high volume work.

Google also pointed to benchmark standing. The company said Gemini 3.8 Flash TTS ranked first on Hume AI's Voice Design Benchmark with a score of 71.4. Voice design, meaning the ability to generate a new voice from a text description, has become a key competitive feature as developers look for distinctive brand voices without hiring voice actors for every project.

The most sensitive capability is voice replication. Google said replication requires a verbal consent recording from the voice owner, and that voice replication through AI Studio is not available in Illinois, Texas, the EEA, the UK, Switzerland or India. That geographic list reads as a map of the jurisdictions where rules on biometric and personal data are most demanding, and it shows Google choosing to wall off the feature rather than test those rules.

The consent design is worth noting for any company considering synthetic voice. Requiring the voice owner's own recorded consent is a meaningful safeguard against impersonation, but it does not eliminate risk. A voice clone produced legitimately can still be misused later, and companies that deploy cloned voices of executives, spokespeople or customer service personas will need policies for how those voices are stored, who can generate audio with them, and how the output is labeled.

The competitive context is crowded. Specialist voice companies, other frontier labs and open source projects are all racing to deliver natural, controllable speech. Google's advantages are distribution, through its developer platform and consumer products, and bundling with the broader Gemini ecosystem. For enterprises, the result is falling prices and rising quality, which makes voice interfaces practical in places they were not a year ago.

The practical question for leaders is no longer whether synthetic voice is good enough. It is whether governance is ready. Voice is identity, and the organizations that set clear rules on consent, disclosure and access now will be the ones able to use these tools confidently when customers and regulators start asking questions.

AI Research Story 6 of 12

Claude Agents Find a New CRISPR Like Enzyme System in Bacteriophages

Anthropic says a swarm of its AI agents has found something biologists had not described before, and the way it got there may matter as much as the discovery.

Anthropic said on September 23 that Claude agents discovered a previously undescribed enzyme system in bacteriophages, the viruses that infect bacteria. The company calls it array associated reverse transcriptases, and the system features CRISPR like DNA repeats. CRISPR systems, originally a bacterial immune defense, became the foundation of modern gene editing, so any new system with similar architecture draws immediate scientific interest.

The scale of the search is the headline for anyone thinking about AI in research. Anthropic said about 950 Claude agents searched for 21 hours using 210 million tokens in the effort that found the system. The agents gathered more than 200,000 reverse transcriptases, identified about 3,500 new candidate systems and narrowed them to 20 for deeper investigation, with human scientists carrying out the laboratory experiments.

That division of labor is the real story. The agents did what humans find slow and tedious: scanning enormous sequence databases, applying consistent criteria, and filtering hundreds of thousands of possibilities down to a short list. Humans did what agents cannot: running experiments in a wet lab and judging whether the results mean what they appear to mean. The combination compresses a search that might have taken a research group months into less than a day of compute.

Caution is warranted. A newly described system is a hypothesis about biology until independent groups reproduce and extend the findings. Whether array associated reverse transcriptases turn out to be a useful tool, a biological curiosity or something in between will take time to establish, and peer review is the test that matters. Companies announcing scientific discoveries also have obvious commercial incentives, which is why the evidence trail matters.

For executives, the lesson extends well beyond biology. The pattern of many parallel agents, a large but structured search space, clear filtering criteria, and humans at the verification step applies to drug discovery, materials science, chip design, legal review and financial analysis. The limiting factor is shifting from how many experts an organization can hire to how well it can frame problems for agents and verify what comes back.

The cost profile is also striking. Hundreds of millions of tokens is a meaningful compute bill, but it is small compared with the salaries of a research team working for months. As agent orchestration matures, expect research intensive industries to restructure around this model, with smaller expert teams directing much larger fleets of AI workers.

AI Models Story 7 of 12

Black Forest Labs Releases FLUX 3 Action, an Open Weight Model for Robot Control

A company best known for generating images has made a serious move into robotics, and it is giving the model away.

Black Forest Labs released FLUX 3 Action on September 23, describing it as a 7 billion parameter open weight world action model for robot control. A world action model predicts how a scene will evolve and what actions a robot should take next, combining the visual understanding developed for image and video generation with the control problem of moving a physical machine.

The company is claiming leadership on a demanding benchmark. Black Forest Labs says FLUX 3 Action ranks first on the RoboLab-120 robotics benchmark, ahead of NVIDIA's Cosmos 3 Nano. It also reported real world results: the company said FLUX 3 Action succeeded in 28 of 30 attempts on a real Franka robot arm, a widely used research platform. Simulated benchmarks and real hardware tests measure different things, and the pairing is meant to show that the model's gains are not confined to simulation.

The strategic logic is interesting. Black Forest Labs built its reputation on image generation models, and the skills involved, learning how the visual world looks and changes, transfer naturally to robotics, where a machine must anticipate the consequences of its movements. Releasing open weights lowers the barrier for research labs, startups and hobbyists to build on the model, which can quickly establish it as a default starting point and create a community around the company's technology.

It also sets up a competitive dynamic with NVIDIA, whose Cosmos family is a central part of its strategy to supply the software foundation for physical AI. NVIDIA benefits when robotics developers build on its models and train on its hardware. A strong open alternative from an independent lab gives developers more choice and puts pressure on every vendor to keep improving.

For enterprise leaders, the takeaway is that robotics AI is following the path that language models took. Capable foundation models are becoming available openly, benchmark leadership changes hands quickly, and the cost of experimentation is falling. Companies in manufacturing, logistics and warehousing that have waited for robotics to mature should start small pilots now, because the gap between research demos and deployable systems is narrowing faster than hardware procurement cycles.

The caveats remain significant. Benchmarks capture narrow tasks, real factories are messier than labs, and licensing terms for commercial use deserve careful review before any production deployment. But the direction is clear: the tools for teaching machines to act in the physical world are becoming a shared resource rather than a proprietary moat.

AI Infrastructure Story 8 of 12

Akamai Signs an $11.6 Billion Cloud Deal With Anthropic and Hands It a Stake

Anthropic's hunger for computing capacity has produced one of the more unusual infrastructure deals of the AI boom, with a company better known for content delivery than hyperscale cloud at the center of it.

Akamai announced on September 24 a seven year agreement worth $11.6 billion to support Anthropic's CPU workloads on Akamai Cloud, with an option to expand by up to $9 billion. The focus on CPU capacity rather than the GPUs that dominate AI headlines is telling. Training and serving large models requires accelerators, but running agents, tools, data pipelines and orchestration at scale requires enormous amounts of conventional compute, often distributed close to users.

The deal also carries an equity component. Akamai issued Anthropic a warrant for about 7.7 million shares, roughly 5% of its common stock, at an exercise price of $111.33 per share, with about 2% vesting at announcement. Tying equity to a customer contract aligns incentives: Anthropic benefits if Akamai's value rises, and Akamai secures a marquee anchor tenant for its cloud ambitions. Similar structures have appeared across the AI supply chain as suppliers compete for the handful of customers able to commit tens of billions of dollars.

Akamai is spending heavily to deliver. The company said it expects about $5.5 billion in capital spending to support the Anthropic commitment, raised its 2026 capex estimate by about $1.7 billion, and left its 2026 revenue guidance unchanged. That combination tells investors the spending comes first and the revenue follows, a pattern familiar from every major AI infrastructure deal.

For Akamai, the agreement is a transformational bet. The company has spent years trying to expand from content delivery and security into cloud computing, competing against giants with far larger footprints. A multiyear contract of this size provides the demand certainty needed to justify major buildout, and it gives the company a credible story as a distributed alternative to the largest clouds.

For Anthropic, the deal diversifies its infrastructure base. Relying on a small number of hyperscalers concentrates risk and bargaining power. Adding a distributed provider for CPU heavy workloads spreads that risk and may suit agent workloads that benefit from running near end users.

For executives buying AI services, the takeaway is that the infrastructure layer beneath the models is becoming more varied and more financially entangled. Supplier relationships now include equity, multiyear commitments and capacity guarantees. Understanding which clouds and facilities sit behind the AI services a company depends on is becoming part of basic vendor risk management.

Funding & Investment Story 9 of 12

Island Raises $400 Million at a $6.4 Billion Valuation to Govern People and AI Agents at Work

Investors are betting that the browser, long treated as a commodity, is becoming the control point for how both employees and AI agents touch corporate systems.

Island raised $400 million in a Series F led by Evolution Equity Partners at a $6.4 billion valuation, the company said on September 24. Sequoia Capital, Coatue, Insight Partners and J.P. Morgan Growth Equity Partners participated in the round. Island says it has about 1,000 employees and has doubled its annual recurring revenue every fiscal year since its 2022 launch.

Island's core product is an enterprise browser, a browser built for corporate use with security, data loss prevention and access controls embedded directly in it rather than bolted on through separate tools. The pitch has always been that most work now happens in web applications, so the browser is the natural place to enforce policy. The company is now extending that logic to AI agents, positioning itself as a layer that governs what both humans and software agents can see and do across enterprise applications.

That extension explains much of the investor appetite. As companies deploy agents that log into systems, read documents and take actions on behalf of employees, security teams face a new problem. Traditional identity and access tools were designed for people clicking through interfaces, not for autonomous software operating at machine speed. A control layer that can observe and restrict agent behavior in the same place it governs human behavior is an attractive proposition for chief information security officers wrestling with agent risk.

The round also reflects broader trends in security funding. Investors are rewarding companies that can credibly claim a role in AI governance, and the largest rounds are going to firms with proven revenue growth and enterprise customer bases rather than early stage concepts. A late stage round of this size gives Island resources to expand internationally, deepen its product and potentially pursue acquisitions.

For executives, the key question is architectural. Where should the controls on AI agents live? Options include the identity provider, the network, the application itself, the model provider, or an intermediary layer like the browser. Each has strengths and blind spots. Many organizations will end up with overlapping controls, and the vendors are racing to become the primary one.

The practical step for leaders is to inventory where agents are already operating inside the business, often through tools employees adopted on their own, and to decide who owns agent governance before a security incident forces the decision. Funding rounds like this one suggest the market for that governance is about to get crowded.

Industry Dynamics Story 10 of 12

Oracle Sends a Force Majeure Notice on Its New Mexico Stargate Data Center

One of the flagship projects of the AI infrastructure boom has run into the constraint that increasingly defines it: power.

Oracle sent a force majeure notice to a Blue Owl Capital unit developing Project Jupiter, its Stargate data center campus in New Mexico, according to reports on September 24. The notice would allow Oracle to delay payments if the facility misses its 2028 target to come online. A force majeure notice invokes contract language covering events outside a party's control, and filing one is a way to preserve rights if delays materialize, rather than an admission that the project has failed.

The underlying problem is energy supply. An Energy Transfer pipeline intended to deliver gas to the Project Jupiter site has been delayed nearly six months, to February 1, 2027, after regulators repeatedly denied permits, according to reports. The campus plan relies on gas to feed on site power generation, so a delayed pipeline threatens the timeline for the entire facility.

Both companies sought to calm the reaction. Oracle said Project Jupiter remains on its planned schedule, and Blue Owl said the notice does not change the financial commitments to the project. Those statements are consistent with a precautionary legal step, but markets tend to read force majeure filings as a signal that at least one party sees material risk ahead.

The episode captures the central tension of the AI buildout. Demand for computing capacity remains enormous, and companies have committed hundreds of billions of dollars to new campuses. But data centers of this scale require vast amounts of power, and the permits, pipelines, transmission lines and generation capacity needed to deliver it move on regulatory and construction timelines measured in years. Regulatory resistance adds further uncertainty, as the repeated permit denials in New Mexico illustrate.

Financing structures amplify the stakes. Many AI campuses are developed by investment firms and leased to technology companies under long term agreements. When delays arise, the question of who bears the cost, the developer, the tenant or the lenders, becomes a matter of contract language, and force majeure clauses move to the center of the dispute.

For executives planning AI strategies, the lesson is to treat compute availability as a supply chain risk rather than a given. Capacity announced is not capacity delivered. Organizations that depend on specific cloud regions or reserved capacity should ask their providers about power sourcing and construction timelines, and build contingency plans for delays that no amount of capital can immediately fix.

AI Infrastructure Story 11 of 12

Microsoft Commits More Than $10 Billion to Cloud and AI Across the Gulf

Microsoft is doubling down on the Gulf states as a hub for AI infrastructure, signaling that geopolitical turbulence in the region has not changed its long term calculations.

Microsoft announced this week that it will invest more than $10 billion in capital and operating costs across Kuwait, Qatar, Saudi Arabia and the UAE through 2030. The company said it plans to invest more than $400 million in subsea and land based connectivity in the region by 2030. Microsoft also said the framework builds on existing commitments to help skill more than 4.2 million people across the four countries by 2030.

The package reflects three strategic priorities. The first is capacity. Gulf governments have made AI a centerpiece of their economic diversification strategies and want local data centers that keep sensitive data inside national borders. Hyperscalers that build locally win government and enterprise contracts that remote capacity cannot. The second is connectivity. Subsea cables and terrestrial links determine how resilient a region's digital infrastructure is, and investment there suggests Microsoft is thinking about redundancy and continuity as much as raw compute. The third is talent, because data centers are only valuable if local organizations have the skills to build on them.

The timing is notable. The region has faced significant instability, and some observers questioned whether the Gulf's ambitions as a technology hub would survive heightened security risks. A major commitment from one of the world's largest cloud providers is a vote of confidence and a competitive move. Countries that anchor hyperscaler investment gain leverage in attracting AI companies, startups and talent.

The Gulf has also become a focal point for the global politics of AI. Access to advanced chips, partnerships between American technology companies and state backed regional champions, and questions about data sovereignty all converge there. Microsoft's investment will be shaped by export control policy and diplomatic relationships as much as by commercial demand.

For multinational executives, the Gulf buildout matters in several ways. It expands the options for serving customers in the Middle East, Africa and South Asia from local infrastructure. It creates new regulatory regimes around data residency that companies will need to navigate. And it reinforces that AI infrastructure is now a tool of national strategy, with governments competing to host capacity much as they once competed for factories.

The broader pattern is clear. Cloud providers are placing long term bets across multiple regions to hedge against concentration risk, and governments are offering incentives to win them. The map of where AI runs is being redrawn country by country.

Enterprise AI Story 12 of 12

Amazon Lets Sellers Run Their Stores Through Claude With a New Seller Assistant Plugin

Amazon has opened one of its most important business tools to an outside AI model, a notable shift for a company that usually prefers customers inside its own interfaces.

Amazon launched a beta plugin on September 23 that lets US sellers use Seller Assistant through Anthropic's Claude or Amazon Quick. Seller Assistant is Amazon's AI tool for merchants. The plugin means sellers can work with those capabilities from inside a general purpose AI assistant rather than logging into Amazon's seller portal for every task.

Amazon's own figures explain the move. The company said Seller Assistant has reached more than 90% of its sellers and that 90% of sellers already use outside AI tools. In other words, merchants are already running parts of their businesses through third party assistants. Amazon can either fight that behavior or meet sellers where they work, and it has chosen the latter. Amazon is also offering primary seller account holders a free 12 month Amazon Quick Plus subscription with sign up open through December 31, 2026, a clear effort to pull sellers toward its own assistant alongside Claude.

The strategic implications reach beyond ecommerce. For years, platforms competed to keep users inside their own dashboards, where they controlled the experience and the data. AI assistants are eroding that model. When a user can ask one assistant to check sales, adjust prices, draft a listing and review supplier emails, the individual portals behind those tasks become back end services. Platforms that refuse to connect risk being bypassed; platforms that connect risk becoming invisible utilities.

For Anthropic, the integration is a significant distribution win. Being chosen as the model through which Amazon's sellers can reach a core business tool places Claude inside daily commercial workflows for a very large merchant base. It also reinforces Claude's positioning as a work assistant that connects to business systems rather than a standalone chatbot.

For executives, the lesson is about interfaces. If your customers, partners or suppliers increasingly work through AI assistants, your products need to be accessible to those assistants through secure, well documented connections. That means thinking about which actions an outside agent should be allowed to take, how to authenticate it, and how to audit what it does.

The companies that design for agent access early will shape the standards. Those that wait may find their customers have quietly moved their daily work somewhere else, with the old portal reduced to a login page nobody visits.