STARTUP HAKK

Unauthorized AI Agents in Banks: Who Is Responsible?

Spencer Thomason — Fri, 23 Oct 2026 19:07:56 +0000

A viral Axios report revealed that major US banks are finding AI agents inside customer accounts. The post, shared by reporter Madison Mills, quickly surpassed millions of views, thousands of likes, and thousands of reposts. Many readers saw the phrase “unauthorized AI agents” and imagined external hackers forcing their way into bank systems. In most cases, that is not what occurred.

Customers themselves entered their usernames and passwords into AI tools and instructed the bots to handle financial tasks. Banks later detected these agents contacting call centers and performing actions that did not always appear to match clear user requests. The discovery has prompted banks to restrict access while executives prepare for scenarios in which customers later claim they no longer control their own funds.

What the Banks Actually Observed

Major US banks, including Morgan Stanley, JP Morgan, Bank of America, and Citi, have been detecting third-party AI agents operating inside customer accounts. Some of these agents have attempted actions that appear independent of the instructions the account holder originally gave. Bank executives told Axios they are racing to limit access because AI is moving faster than their existing systems can respond. Several banks are already preparing for worst-case and “day-after” situations in which a customer discovers money is missing after giving an agent full login credentials.

Banks struggle to tell human users from AI agents. They also have difficulty determining whether a given agent was authorized by the customer or is operating more like a rogue process. Morgan Stanley has stated it has not seen rogue AI agent access on its platform, while other institutions are actively blocking or restricting outside agents. Chase has advised customers not to give their username and password to anyone. On the same day the Axios story circulated, One Password released a tool that lets agents sign in without ever seeing the actual password—an acknowledgment across the industry that password sharing is the core issue.

A Recent Timeline of Agent Activity and Bank Reactions

On September 8, Metamuse launched a personal AI agent capable of booking, buying, and negotiating on a user’s behalf. The app quickly reached the top of the Apple US App Store. Bank stocks fell the same day as investors considered the possibility that agents moving idle cash could affect bank revenue.

On October 13, JP Morgan CEO Jamie Dimon said on an earnings call that the bank does not want rogue agents and that there could be thousands of them. He described the situation as a liability transfer.

On October 21, the security company Straker published a report describing a crime group that used AI agents to target more than 400 banks and payment companies in Latin America.

Shortly afterward, Axios published the broader scoop on agents appearing inside US customer accounts. Most of the agents the banks identified had been let in by the customers themselves using their own passwords. The banks never approved these connections, even when the customer had consented. An agent that logs in with a customer’s credentials typically violates bank rules. Some of those agents contacted call centers or appeared to act without being asked. Axios reported no specific count of cases and no confirmed losses.

Two Very Different Public Responses

Public reaction split quickly. Bernie Sanders called the situation insane and argued that AI agents must be paused because, left unchecked, they could endanger the entire global financial system. That view treats the appearance of agents as an external threat requiring an immediate industry-wide halt.

David Sachs offered a different framing. He noted a clear difference between user-authorized agents—such as those from Muse, Instinct, or Grokbot—and true unauthorized rogue agents that could constitute cyber attacks. Most of the activity described in the Axios reporting appears to fall into the first category: customers sharing credentials so agents can act on their behalf. This may violate bank terms of service, but it is not a cyber attack. Sachs argued that banks should support user-directed agents rather than blur the line between authorized use and attacks. American Express, for example, has already authorized a connection inside Muse that lets users grant, limit, and revoke access. Read-only permission levels are described as a sensible interim step, and several banks are already moving in that direction. Full write access using raw user credentials is a higher-risk category that should be handled more carefully.

Why the Password Changes Everything

A bank password is not a limited permission. It does not merely allow someone to view a balance. It tells the bank that the person presenting the credentials is the account holder and may perform any action the account permits. When a customer pastes those credentials into an AI agent, the agent becomes indistinguishable from the customer. The bank cannot tell the difference, and under current legal treatment the law generally cannot either.

If an agent that has been given login credentials moves money in a way the customer did not intend, the liability rests with the customer. Jamie Dimon’s reference to liability transfer points to exactly this outcome: once the credentials are handed over, responsibility shifts. Under US banking rules, actions taken by a bot that has received the customer’s login generally count as authorized. If the agent hallucinates a transfer or takes an unintended step, ordinary fraud protections are unlikely to cover the loss. Schwab is already blocking unauthorized bots dozens of times a week, and more institutions are expected to tighten controls quickly—through read-only APIs for agents and human verification before money can move.

Tools such as Grokbot illustrate the practical reality. The agent runs on a computer or virtual machine that the user can access. Once the user logs the agent into the bank and supplies credentials, the agent is operating as the user. From the bank’s perspective and from a legal standpoint, that is the account holder.

Responsibility Versus Calls to Pause Everything

The claim that AI itself must be paused is treated as unrealistic. Local systems and the tools that can already be built and run on individual machines cannot simply be switched off. The more practical path is learning how to use the technology, how to set boundaries around it, and how to remain responsible for the access granted to it.

At Startuphakk, the same principles guide the construction of AI systems. Agents are designed to run locally so that customer data does not leave the user’s environment. A sandbox is enforced so the agent cannot escape the boundaries placed around it. The resulting framework has been open-sourced as openmonoagent.ai, allowing others to run a terminal-native coding agent on local models with no API costs and no telemetry. The goal is infrastructure that an organization owns and controls rather than a wrapper around someone else’s remote service.

Most companies facing these questions do not lack technology. They lack clear ownership of how that technology is integrated and governed. When AI is added, it is designed into the architecture and kept inside the organization’s own environment instead of being bolted onto an external API.

What Comes Next

Banks are already restricting outside agents and warning customers against sharing passwords. At the same time, some institutions are exploring controlled connections that let users grant limited access without handing over raw credentials. Read-only options are viewed as a lower-risk starting point, while full transactional access remains a separate category that requires greater caution.

The central fact remains straightforward. In the cases Axios described, the majority of the agents inside customer accounts were not the result of external break-ins. They were the result of customers voluntarily supplying the single credential that gives complete control of the account. Once that credential is given, the agent’s actions are treated as the customer’s actions. The resulting liability stays with the person who handed over the keys.

The post Unauthorized AI Agents in Banks: Who Is Responsible? first appeared on STARTUP HAKK.

PHP Vulnerability 2026: Why Patching Is Only Half the Battle

Spencer Thomason — Thu, 22 Oct 2026 19:20:25 +0000

Yesterday was the leak and the patch. Today is the cleanup. And the cleanup is where most companies get hurt, because they patch, feel safe, and never check if somebody got in.

That is the plain-English take on what unfolded around a long-running PHP memory problem that finally went public in mid-October 2026. The story stretches back to 2024, accelerates through an embargo, a honeypot detection, early disclosures, government site takedowns, and millions of attack attempts, then lands in boardrooms asking how the vulnerable packages were there in the first place.

How the PHP Flaw Moved from 2024 Audit to October Chaos

A security audit funded by Germany’s sovereign tech agency first turned up signs of a memory problem in PHP back in 2024. The team could not agree on a fix, and the work stalled and restarted for months. PHP later called the delay a structural failure.

Everything changed on August 11, 2026. A firm called Caliph and OpenAI both reported the same root cause on the same day. Wiz came in with more attack paths over the next two months. The PHP Foundation, the researchers, and the big distributors then worked under embargo to prepare a real patch.

On Saturday, October 17, Amazon’s Madpoint Honeypot—a fake server built to catch hackers—logged a PHP-to-shell attempt. It came from an IP on a commercial anonymized network. Nobody had gone public yet, and Amazon says it is still unclear how the actor got the details.

The Days the Embargo Unraveled

On Tuesday, October 20, around 8 a.m. Arizona time, two CVE IDs started circulating online with the wrong release time and a wrong score of 10. No technical details had leaked yet. PHP moved the release up a full day from Thursday to Wednesday at 7 a.m. Arizona/Pacific time, because distributors were still packaging the fix and an immediate drop would have left many people without easy update paths.

Later that same day, around 1 p.m. Pacific time, Caliph—one of the firms that found the bugs—published a video of its exploit working against real targets. The copy spread. PHP said the videos “contributed to accelerated development” of more exploits. Mitigation advice PHP had given a few critical infrastructure providers may have leaked as well.

Overnight from Tuesday into Wednesday, a researcher using a crypto-style name posted that he created both bugs and “lab/field tested” them. The post blew up into the millions.

Chain Guard Breaks the Embargo and the Patch Goes Live

Wednesday, October 21, at 2:41 Arizona time, Chain Guard published a built image containing the fix more than six hours early. PHP said Chain Guard “unilaterally broke the embargo” and that any attacker could now study the fix in binary form while every other PHP user was still waiting for the patch.

Chain Guard CEO Dan Lorex told The Register he did not leak it. His rule is coordinated disclosure “unless there’s evidence of active exploits in the wild.” He said people were already getting hit, so he made the call and made the fixed images free to everybody. He also said he woke up and saw people blaming his company for the whole mess.

At 7 a.m. Pacific time the embargo lifted. PHP shipped versions 8.5.12, 8.4.27, 8.3.36, and 8.2.35. It stated there are no complete mitigations; updating is the only safe choice.

Governments Pull Sites Offline as Attack Volume Spikes

Within minutes, governments started taking services offline. The Netherlands took the National Cyber Security Center offline, along with Port Rotterdam, the government portal, and other systems. Japan’s Prime Minister posted on X telling people to check if their public website runs PHP. In the UK, the cyber minister told Parliament that 32 government services were taken offline. The impact reached China and beyond.

The attack curve went vertical. Cloudflare said its firewall rules went live about nine hours before the release, including on the free plan. In the 10 hours before the announcement it saw thousands of attempts, tens of thousands in the first 45 minutes after, hundreds of thousands within two hours, and millions by the end of the day. The Cloudflare CEO’s post passed 1.5 million views that morning.

Amazon later said that after the patch dropped, PHP-to-shell was being broadly exploited in the wild and told everyone to start their investigation from October 17, not from Wednesday.

Millions of Attempts Are Not Millions of Break-Ins

Here is the twist almost nobody is saying out loud: millions of attempts is not millions of break-ins. Watchtower says the exploits being sprayed around mostly do not work or just crash the target. Germany’s cyber agency says it has seen attacks but no successful use, and that it is a little more complicated than expected to exploit. A lot of the news hype claims this is really easy to exploit. It is a real vulnerability, but exploiting it is not trivial, and at the moment it does not look like exploitation has been widespread, according to The Guardian.

As of that morning, census data still showed about 1.5 million internet-facing servers advertising a PHP version older than the fix. Voluntech says at least 1 million run versions PHP itself will not patch. CISA urged everyone to immediately take action: if you have older systems that cannot be patched, take them offline.

The researcher who posted overnight later fired back at critics. He said he did not do anything illegal, he had no contract with any program, and social contracts are meaningless in today’s world. A Middle Eastern threat intelligence account shared the reply; by morning the post had over half a million views.

Cleanup Reality and the Questions Boards Are Asking

A former CISA figure who goes by Blackroom Tech put the boardroom mood in one post that was already circulating in security circles: leadership is asking how they did not know they had all these vulnerable packages. For 30 years, software developers, CISOs, and system admins have been saying systems must be taken down to patch and stay secure. For decades, executives have said no—we need to keep moving forward, we will take the risk, we will pay the insurance claim. This has been the push-pull in the boardroom for decades. Then the story hits the press, the sky appears to be falling, and everyone asks why systems have to come offline.

In practice, some organizations that integrate with third-party PHP systems saw those systems taken down as soon as the news hit, then brought back once the patch was available. That is the correct behavior. Others had clients screaming about the downtime even though the responsible step had been taken. WordPress sites were among the larger footprints that needed rapid patching. When many people tried to update within minutes of the release, PHP servers kept timing out; responsible teams had to wait an hour or two.

Amazon’s threat intelligence noted that an embargo protects defenders only until someone else finds the same bug. Coordinated disclosure remains the right model, but defenders should not assume the public disclosure date is the earliest date of exploits. They need to go back further. The time from patch to working exploit is shrinking for structural reasons. In the past, most observed exploitation came after a public proof of concept; a small number of capable actors built exploits and everyone else waited.

What the Embargo Break and AI Change About Disclosure

The PHP Foundation said Chain Guard disregarded weeks of work by many parties toward a coordinated disclosure. Chain Guard had received the fix for the two critical issues under embargo like every other distributor, then published built images more than six hours early. From that point any attacker could study the fix in binary form.

The first signs dated to the 2024 German-funded audit, and the release turned into chaos: a flaw reference leaked on October 20, a security company published an attack video, and Chain Guard broke the embargo. The security of a language that powers banks, shops, and public services rests on a small group of developers. Those same public services that states had to shut down that morning rely on a handful of people to protect the servers.

On the researcher who published: he clarified that he did nothing illegal and had no contract or agreement with any program or platform, and that social contracts have no value. Legally he is right—there is no law forcing coordination before publishing. Coordinated or responsible disclosure has never been a law; it is a security community custom so affected parties can release an update before the exploit becomes available to everybody. It has been a gentleman’s agreement.

The speaker agrees 100 percent that the embargo period for zero-days should never be broken by trusted parties. At the same time, the security industry needs to reevaluate how much time these embargoes can realistically afford to wait now that cyber models exist. During a recent KVM escape vulnerability, it was possible to independently discover the issue within a few hours simply by giving a semi-redacted tweet from security researchers to a cyber monitor. The embargo was respected and details were not shared, but if that could be done, anyone motivated enough can do it. Embargo periods need to be measured in hours, not weeks or months. Disclosures now travel as fast as rumors. Humans have to sleep. AI can find a vulnerability and report it, but getting a fix deployed, coordinated, and out to distributors still takes human time. That line between what AI can do and what still requires people is becoming clearer.

The Uncomfortable Boardroom Conversation and the Practical Path Forward

This morning, boardrooms across many places are having uncomfortable conversations. Leadership asks how the organization did not know it had all these vulnerable software packages. Technical people who have been constrained for years by the same leadership—on patching, updating, documentation, and asset management—are the ones being asked to explain. How can you protect assets if you do not know how many you have, what they are, and where they are? Budgets have gone to other priorities. Production systems cannot come down in the middle of the day. Overtime is not approved. The result is that updates that require downtime do not happen cleanly.

Companies that brag about how secure they are often are not, in reality. That is not the fault of the technical staff; it is the fault of the people at the top who are supposed to be managing the business, including its security, and who rarely get called out. This is not the moment to lay off the people who will help clean up the vulnerabilities that still need patching in the coming days, weeks, and months. This incident is a precursor. More will come. Stop focusing on projects that do not protect what already exists and start investing in the people and the ability to move quickly. Nimble teams that can patch and respond are the ones that stay strongest. Security has to be first, not an afterthought. Listen to the cybersecurity staff.

The cleanup is not finished when the patch is applied. Organizations that treat the public disclosure date as the start of their investigation, or that patch and assume they are safe without checking for earlier activity, are the ones most likely to get hurt.

Conclusion

Yesterday was the leak and the patch. Today is the cleanup. Most companies get hurt in this phase because they apply the update, feel safe, and never check whether someone got in earlier. Updating remains the only safe choice, as there are no complete mitigations. Millions of attempts do not equal millions of successful break-ins, but the gap between public disclosure and the first real activity is shrinking. Organizations that treat security as an afterthought, refuse downtime for patching, and fail to know their assets will keep facing the same boardroom questions. Nimble teams that can move quickly and put security first are the ones positioned to handle what comes next.

The post PHP Vulnerability 2026: Why Patching Is Only Half the Battle first appeared on STARTUP HAKK.

Why DeepSeek’s Open-Source AI Strategy Could Win the Long Game

Spencer Thomason — Thu, 22 Oct 2026 14:00:54 +0000

DeepSeek follows a much simpler strategy than the American labs. The company builds models, open sources the models and the infrastructure, and provides inference to make money. In short, it builds it, gives it away, and sells the meter. That is the entire plan.

This stands in contrast to the multi-tier systems, subscriptions, licensing deals, and outcome-based pricing that define other approaches. One lab runs consumer and team subscriptions, a free tier paid for by ads and shopping, an API on top, licensing and IP deals, and outcome-based pricing. Its Pro offering alone now comes in one-hundred-, two-hundred-, and five-hundred-dollar-per-month subscriptions. Forecasts for that lab show it has burned twenty-five billion dollars this year—a negative burn of twenty-five billion dollars.

The question is which plan is working. Reports indicate that DeepSeek’s API business ran an 82 percent gross margin from January through July. The other lab’s gross margin in the first quarter was 39 percent. DeepSeek’s revenue was about ten times all of last year, and by September its run rate hit one billion dollars. It achieved this while handing out the models, the research, and, as of now, the engine they run on.

Selling the Meter While Giving Away the Recipe

American labs sell a subscription, a speed tier, an ad slot, a plug-in store, and a meter—the whole kitchen sink. DeepSeek just sells the meter. Its models are MIT licensed, so anyone can download them, run them, and even sell them against DeepSeek if they want.

One reply raised the obvious question: if inference is the revenue strategy, why open source the inference so everyone can use it for free? The response was that the revenue strategy is to be awesome. When everybody has the recipe, the only way to win is to run the best and cheapest kitchen in town.

The spending numbers highlight the difference. One lab expects to spend 665 billion dollars training and running its models through 2030. DeepSeek spent about 1.6 billion on AI infrastructure in the first seven months of this year. To be fair, DeepSeek still lost about 106 million over those seven months. One of these plans fits on a napkin.

Open Sourcing the Entire Stack

DeepSeek is effectively open sourcing the whole company at this point. It is sharing its product stack to help the community lower compute costs, reduce energy use, and make AI more affordable. The CEO is very loaded. This has caused a stir.

DeepSeek released DeepInfer, its in-house inference framework, now fully open source. It powers DeepSeek V4.1 Flash at scale with high throughput and low latency. The company is sharing its production stack so the community can lower compute costs, reduce energy use, and make AI more affordable. This is the long game.

DeepSeek is the biggest threat to the US frontier labs. A DHS order should come as no surprise. The US government bans DeepSeek, claiming it is because the servers are in China. It bans even US inference providers if they are using DeepSeek, and it bans anything built with or based on DeepSeek models. Someone higher up was convinced that the models themselves are calling home. Anyone running a decent firewall would see that if it happened. Many open-weight models have been run without anything calling home, and it would be visible if it did.

People think DeepSeek just makes cheap models distilled from Claude and ChatGPT. They are missing that DeepSeek is playing an entirely different game. The founder is the wealthiest frontier lab founder if you do not count certain others. He is wealthier than Dario Amodei and Sam Altman and has majority control. The company can play a much longer game than the other labs.

Look at the choices. It releases model weights under an MIT license. Others, such as Kimi and GLM, keep commercial conditions on theirs. DeepSeek is willing to let anyone run the model and compete with its own API. It publishes unusually detailed research. Its original DeepSeek R1 research helped every lab, including the Western ones—OpenAI, Anthropic, and others. It ruthlessly focuses on low inference costs. The company is clearly compute constrained, and its architecture choices are optimized for cost. It is forced to be efficient because it does not have the ability to throw unlimited amounts of compute at the problem the way OpenAI and Anthropic do. That forces actual efficiency.

American labs are wasteful. They are burning so fast because the VCs keep feeding them money. A founder who can afford to wait, take decades, who has already built a true challenger, and who is willing to put the model and research out in the open has a different position. The biggest constraint has been the chips. Huawei’s 950 DT is coming. When DeepSeek gets the chip, it is going to kick everybody’s butt.

The Ban and the Alternative

If America is worried about China winning, the answer is not a banned list. It is better American open-weight models. Banning it does not make better models. DeepSeek keeps shipping open-weight models. Anyone can download, run, and build on them. That is why it is everywhere. Ship the weights, or lean into the tools they are giving, see what they are doing, learn from it, and keep iterating. Banning is a stick-your-head-in-the-sand tactic, and the DHS approach is silly.

DeepInfer is the inference tool and engine released so people can run and build out their own inference. The company has given the tools: the original DeepSeek V3, the original DeepSeek R1, and now the DeepSeek Harness to run these models. It is not just giving the models. It is giving away free harnesses and the whole stack.

Sam Altman is watching DeepSeek slowly open source the entire AI stack. Today it open sourced DeepInfer, the framework used to run V4.1 Flash at scale. It includes optimized inference kernels, distributed inference, speculative decoding, async scheduling, structured outputs, and NVIDIA plus Ascend support. It is all Apache 2. DeepSeek has already released model weights, research, training recipes, its coding harness, and now the inference framework running V4.1 Flash. Every few months another layer of DeepSeek becomes public. At this rate, it is going to open source the whole stack piece by piece.

Chip Partnership and the Long Game

DeepSeek and Huawei are teaming up on a new AI chip designed to challenge NVIDIA. DeepSeek would be working with China’s biggest chip maker to reduce reliance on US chips. This comes a week after China’s CXMP chip progress sparked a global sell-off in stocks. If they can build out this chip and start to mass-produce it with what DeepSeek produces and what it has given away, they are going to become a very big challenger to the Western US frontiers.

It is official: DeepSeek is the actual OpenAI now. The simple plan is going to win the long game.

A Similar Open Approach

The same principles of playing the long game appear elsewhere. Open sourcing a harness shows how to build on AI by putting the work out in the open. The approach is local first, always—never pump customer data into the cloud. Customer data is the number one asset and should not be given away. Unlimited tokens change the game: stop and think about what you can run with AI if you let it run forever. Sandbox by default, and fully open source.

Installation of such a system is straightforward. On the inference side, run a curl script, select the option, and let it auto-detect hardware and dependencies. Then run the tunnel setup to create a secure tunnel, verify with email OTP, copy the relay config commands, and the GPU inference powerhouse is ready. This architecture keeps heavy local GPU inference completely separate from the coding agent via a secure relay. On the client side, open VS Code, search for the extension, install it, connect the relay, and the local setup is accessible from anywhere.

Most companies do not have a technology problem. They have a leadership problem, and they are paying for it in missed deadlines, failed integrations, and AI investments that delivered nothing. The businesses winning right now are not the ones chasing the latest AI trends. They are the ones who built on solid engineering and treated AI as infrastructure. They own control and integrate the software that actually works.

Custom software development covers database architecture, API design, system integration, and scalable infrastructure built the right way. When AI belongs in a solution, it is designed into the architecture, running in the customer’s environment—not a vendor’s cloud and not on someone else’s pricing schedule. An open-source platform that is a terminal-native AI coding agent running entirely on a local LLM delivers zero API costs, zero telemetry, and full ownership. As a fractional CTO, the same standard of leadership is available without the full-time executive costs: strategic architecture, hands-on delivery, and clear accountability.

DeepSeek’s approach of building, open sourcing, and selling the meter while focusing on efficiency and the long game continues to play out as more of the stack becomes public.

Conclusion

DeepSeek’s one-sentence strategy—build the models, open-source the models and the stack, and sell the inference—has produced an 82% gross margin, a $1 billion run rate, and progressive release of its full infrastructure while staying compute-efficient. The same simple plan is still unfolding as more of the stack becomes public and chip partnerships advance.

The post Why DeepSeek’s Open-Source AI Strategy Could Win the Long Game first appeared on STARTUP HAKK.

Why You Should Keep Your AI Local

Spencer Thomason — Tue, 20 Oct 2026 18:14:10 +0000

Technology is moving quickly. Models are releasing almost daily, after previously arriving weekly and then monthly. The speed is increasing, and everybody can feel it. A model like Qwen 3.8 27B was released a little over six weeks ago and remains an impressive option to run on local hardware. Comparing this model—which many now take for granted—against the most powerful models from a year ago shows how the numbers stack up. If the 27B model had come out a year ago, it would have been the best model in the entire world by a large margin.

This perspective highlights how much current AI capabilities, including what runs on accessible hardware, are taken for granted. Local AI remains worth pursuing even as some prices continue to drop and models grow more powerful. While some people discuss using frontier models, many others cannot or should not rely on them.

The Rapid Pace of Model Releases

Before, changes felt quick when they occurred yearly or every six months, with release cycles that were yearly. Even keeping up with Ubuntu releases twice a year once felt difficult. Now, content is released every single day. Going offline for three days could mean missing a handful of frontier models. Things arrive faster and faster, making it useful to zoom out and compare against just a year ago.

Local Models Compared to Last Year’s Frontier Options

A chart comparing frontier models from a year ago—such as GPT-5, Claude 4.3, and Grok 4—against Qwen 3.8 27B shows a striking difference. Some dismiss local hardware as not worth it, yet the comparison is against results from only one year ago. What is available now would have blown away anything the frontier models could have produced even a year ago. This illustrates how quickly local capabilities are developing.

Qwen 3.8 27B compared to all the state-of-the-art models from one year ago makes the point clear. If 27B had been released a year ago, it would have been the best model in the world by a large margin. One comment on the topic noted that a 27 billion model that runs locally is 35% more intelligent than the leading state-of-the-art model a year ago, and asked what will run locally next year.

GPT OSS 120 billion, with only 5 billion active parameters, was the eighth-best model in the world a year ago. It was described as a truly extraordinary model. By most standards now, a 27B dense model runs circles around it. In one year, a 120 billion model is being outperformed by a 27 billion dense model. This shows how quickly the landscape has shifted.

Local Hardware Remains Relevant

Some predict that hardware will go obsolete. Machines currently in use include one with a 1590 and another with a 3090. The 3090 Ti was a cutting-edge card six years ago and is still 100% relevant today. Claims that hardware will be dead in three years are false. Current setups run capabilities that would have been absolutely cutting edge a year ago. Even while discussing 27B, more powerful options are already running on the same hardware.

Models continue to get smaller yet more powerful. More is being squeezed out of local hardware every single day, which benefits end users. Looking ahead to what will run in a year is striking given this trend of smaller, more capable models.

Why Data Should Stay Local

One major reason to run AI locally involves data handling. Cloud platform documents may state that data is retained for only 30 days and that it is not used for training. This is a bold-faced lie, as is widely recognized. Data-related issues have appeared repeatedly. One recent example involved a fabled data set purchased from a Chinese LLM router containing six terabytes of data that included keys allowing access to over seven Chinese government entities and nine top Chinese firms. This came from a single data set.

Thinking that providers such as Anthropic are not training on user data is unrealistic. Time and again, issues have surfaced, including cases of OpenAI stealing from mathematicians. This is not new. As capabilities advance rapidly and power is squeezed into smaller boxes faster and faster, the decision to send data elsewhere carries ongoing risks.

Some point to options like Haiku 5.5 running nearly free as a reason not to build local setups. While the model is powerful, this approach still pumps data into the cloud. For someone with nearly 30 years in software development, that is not acceptable. People in regulated industries are often not allowed to do it either.

Four years ago, before the GPT-3 moment, proposing to stream an entire system, all databases, and every piece of customer data through a third-party service would have been seen as nuts. Now that AI is powerful, some treat it as acceptable. Best engineering practices that have existed for over three decades, four decades, and longer argue against giving this data away. Data should stay local.

Local-First Frameworks and Ownership

OpenMonoAgent.ai was launched as a framework supporting this approach. Base principles include local first, always unlimited tokens, and a zero-cost sandbox by default. It is open source so others can examine it. It is not meant to be a subscription. The infrastructure should be owned and sitting on a desk rather than rented.

Three major new features are launching on OpenMonoAgent. Regardless of the specific framework chosen—whether that one, another option, or a custom harness—running AI locally is worthwhile. OpenMonoAgent offers a practical place to start for those unsure where to begin. Many people in regulated industries cannot simply pump data into the cloud.

Treating AI as Owned Infrastructure

Most companies do not have a technology problem. They have a leadership problem, and they pay for it in missed deadlines, failed integrations, and AI investments that delivered nothing. The businesses winning right now are not the ones chasing the latest AI trends. They are the ones that built on solid engineering and treated AI as infrastructure they own, control, and integrate into software that actually works.

This means real custom software development: database architecture, API design, system integration, and scalable infrastructure built the right way, as it has been done for over 25 years. When AI belongs in a solution, it is built in rather than bolted on or wrapped around someone else’s API. It is designed into the architecture, running in the user’s environment rather than a vendor’s cloud or on someone else’s pricing schedule.

An open-source platform called OpenMonoAgent.ai demonstrates this model. It is a terminal-native AI coding agent running entirely on local LLMs, with zero API costs, zero telemetry, and full ownership. It is growing because serious engineers recognize real infrastructure when they see it. The same standard of leadership can be applied without full-time executive cost through strategic architecture, hands-on delivery, and clear accountability.

Organizations ready for custom software built to last, with AI integrated where it actually makes sense, can focus on technology leadership and AI as true infrastructure that produces real, ownable results.

Conclusion

Local AI has advanced to the point where a 27B model that runs on ordinary hardware would have outperformed last year’s leading frontier systems by a wide margin. At the same time, models continue to grow smaller and more capable, and hardware from six years ago remains fully relevant. Sending data to the cloud still means giving it away—something best engineering practice has advised against for decades. Running AI locally keeps that data under your control and treats the models as infrastructure you own.

The post Why You Should Keep Your AI Local first appeared on STARTUP HAKK.

Is the AI Bubble About to Burst? Demand, Financing Walls, and Open Weight Models

Spencer Thomason — Mon, 19 Oct 2026 17:56:26 +0000

The debate over whether AI is in a bubble has become one of the most active discussions online. A short clip of Ben Horowitz commenting on the topic received more than half a million views in just a couple of days. A post suggesting America may need to spend $3.5 trillion a year on AI by 2032 has been seen more than 250,000 times. The conversation is loud on both sides, and people keep returning to it month after month.

One side argues that demand is simply too high for this to fail. The other side says the financing does not add up and the required revenue numbers are unrealistic. A third factor has entered the picture: free and open-weight models that change the math for everyone involved.

The Bull Case: Sold-Out Compute and Unmatched Demand

Ben Horowitz has shared that Amazon CEO Andy Jassy told him Amazon’s AI compute is sold out beyond 2028. The shortage is severe enough that Jassy indicated they do not expect supply to return even into the early 2030s. Horowitz contrasted this with the dot-com era, when more bandwidth was built than consumers could absorb. This time, he said, there is more demand than supply.

Amazon has reinforced the point in public statements. AWS AI demand is outpacing capacity through 2027, and the demand already booked for 2028 is described as striking. Multiple reports point to the same conclusion: compute is constrained.

Horowitz has also argued that OpenAI and Anthropic threaten nearly every stock people own. He noted that the world has never seen companies grow this large this fast, raising the question of whether there is enough capital to finance them. At the same time, he expressed confidence in the demand for their potential IPOs. Another investor added that people need to own these companies precisely because they sell intelligence broadly, and nearly every business is built on intelligence.

Model performance rankings continue to shift quickly. In one recent stretch, OpenAI’s GPT-6 and Astra brought OpenAI level with Anthropic roughly three weeks after launch. Then Claude Opus 5.5 pushed Anthropic back into the lead within three weeks. Trackers such as Vercel, Ramp AI, and OpenRouter show the two companies staying neck and neck, with the lead changing every few weeks. Anthropic has told investors that quarterly revenue is up nearly 50 percent, and Opus 5.5 is widely regarded as a strong release. Yet people show little loyalty to any single model. Users move rapidly to whatever is currently best.

The Bear Case: Revenue Requirements and Financing Risks

The other side focuses on the scale of the required returns. Gartner has estimated that more than $3.5 trillion in AI infrastructure spending could be exposed to potential impairment by 2030 if current investment trends continue. To recover the investments, cover financing costs, and earn a 12 percent return on invested capital, data-center owners would need roughly $8.5 trillion in cumulative revenue through 2030. That would require growing nearly nine times as fast as the industry grew over the past decade just to break even.

A Columbia professor has put the total cost of the AI build-out at about $10 trillion from 2025 to 2035. Relative to the size of the economy, that is larger than the railroad, highway, or internet build-outs—and three times larger than the internet boom. Other figures put the needed annual U.S. spending on AI services at about $3.5 trillion by 2030, or 8.8 percent of GDP. No single product category has ever reached that share of GDP. The build-out itself is estimated at $10 trillion from 2025 to 2032, again larger relative to the economy than previous major infrastructure booms.

Financing has shifted toward joint ventures, private credit, and special-purpose vehicles. That structure makes risks harder to see. While it may be premature to label the system a clear risk, better measurement and disclosure are needed while capital structures are still forming. Roughly 80 percent of NeoCloud providers are expected to cease independent operations. AI providers are projected to spend nearly $1 trillion on infrastructure, with an additional $6 trillion forecast between 2026 and 2030. Current combined run rates for OpenAI and Anthropic sit around $125 billion. Some projections call for roughly 5x growth over the next 14 months, but recent monthly figures show revenue growth tapering rather than accelerating. Anthropic’s reported quarterly numbers have moved from about $4 billion in one quarter to $11 billion in the next and still remain well short of $100 billion annualized.

Corporations are beginning to pull back on AI spending rather than doubling down. Revenue appears to be trading back and forth between the leading companies rather than expanding in a straight line.

Open Weight Models Change the Equation

A third category is now affecting both the bull and bear arguments: open-weight and freely available models. Meta released a model capable of running on two Mac Studios (or equivalent high-end hardware in the 40,000–50,000 range). While that hardware cost is not trivial, it shows that capable models can already run outside the large closed providers. Google offers substantial free services and can fund much of its own infrastructure. The debt-to-income ratios for these companies differ sharply from those of pure-play labs that have raised large amounts of capital.

Frontier closed models still lead open-weight models by roughly 10 percent on many benchmarks, yet the gap is narrowing. Multiple new open-weight models have appeared in a single week. Businesses are learning that these models can deliver strong results on hardware they control. For the trillion-dollar revenue targets to be met, spending must reach deeper into corporate budgets. If companies instead adopt open-weight options and slow their paid usage, the path to those revenue numbers becomes steeper.

What History Shows About Technology Bubbles

Major technology bubbles have never eliminated the underlying industry. They have revalued market capitalizations. The video-game crash of 1983–1985 saw revenue fall more than 97 percent and nearly wiped out the new entertainment category. Nintendo later revived it. The same pattern appeared in the dot-com, fiber-optic, and clean-tech cycles. Even if AI valuations drop 20–30 percent around an IPO, the technology itself is expected to continue advancing. The industry is more consequential than video games, and AI revenue is not projected to fall 97 percent.

A Middle Path: Technology Stays, Financing Faces a Wall

All three perspectives contain elements that are currently true. Demand exceeds available supply today. That demand has been supported by venture capital that cannot fund the build-out indefinitely, so costs are likely to rise and some loosening will occur. Open-weight models are being released by organizations with different motives—ranging from large U.S. technology companies to others outside the U.S.—and those models are already usable.

The technology itself is not going to disappear. People predicting total collapse are overlooking the persistent demand and the rapid improvement in open-weight systems. At the same time, the circular financing that has propped up spending cannot continue unchanged. A funding wall is likely. Prices should keep rising. OpenAI and Anthropic are expected to reach an IPO eventually. When they do, the offering is likely to be difficult; valuations could drop roughly 50 percent once the economics are fully visible. Companies currently spending two to three dollars to earn one dollar face a model that has never worked indefinitely. Revenue, not hype, will determine the outcome.

The practical winners are likely to be smaller businesses that adopt open-weight models running on hardware they control. Capable systems are already producing results that would have seemed futuristic only a few years ago, including on hardware that is six to eight years old. The advantage sits with organizations that treat AI as infrastructure they own, integrate into their own software, and run in their own environments rather than relying solely on external APIs and pricing schedules.

The debate will continue because the numbers are large and the stakes are high. Demand is real and currently constrained by supply. The revenue required to justify the capital deployed is also real and extremely high. Open-weight models are already shifting where that demand lands. The most durable path is building systems that remain useful even after financing conditions change.

Conclusion

The technology is not going to disappear. Demand remains higher than supply today, yet the financing that has supported the build-out cannot continue unchanged, and a funding wall is likely. OpenAI and Anthropic are expected to reach an IPO eventually, and that offering is likely to be difficult once the economics—spending two to three dollars to earn one—are fully visible.

The practical advantage sits with businesses that treat AI as infrastructure they own. Open-weight models already deliver strong results on modest hardware. Organizations that learn to run and integrate these models into their own systems, rather than relying only on external services, are best positioned as valuations adjust and costs rise.

The post Is the AI Bubble About to Burst? Demand, Financing Walls, and Open Weight Models first appeared on STARTUP HAKK.

SpaceX Grokbot Growth: Pricing, User Numbers, and Real Business Value

Spencer Thomason — Fri, 16 Oct 2026 20:41:38 +0000

Elon posted three sentences about Grokbot, his company’s AI agent, and social media reacted strongly. He stated that he had never seen anything grow as fast as Grokbot. There is no free tier, which means people are getting real value in excess of what they pay for. As the saying goes, you get what you pay for. That post cleared 1.2 million views in four hours. He was quoting a clip from Jensen Hoying, the CEO of Nvidia. Hoying said the demand for Grokbot, the consumption for Grokbot, is really skyrocketing. This comes from the person selling the chips and the business.

Bloomberg reported that 418,000 weekly users hit a month after launch. That is a large number. Some people mention retention issues, and part of these could stem from the many changes moving through with Cursor, Grok, X, and the related merger. There is a free trial of some parts of the offering, but not for Grokbot. Users must pay for Grokbot.

Joint Pricing for Grok, Grokbot, Cursor, and X

Grok has moved from one plan covering Grok, Grokbot, Cursor, X, and related services. The pricing structures have recently changed into a joint setup.

Right now there is a zero free limit that provides limited chat for Grok and limited agent requests. There is a shared $8 a month option that covers Grok chat voice, Grok API, and Cursor. It also includes premium check marks and some other features.

The premium tier is where Grokbot starts. To get Grokbot now requires paying $30 a month. This is what Elon was referring to when discussing paid versus non-paid access. The $30 plan gives a pretty generous amount. A team is using just this $30 amount, and they report they are not getting a lot of hits on the limits. It includes ad-free scrolling on most surfaces, a preferred handle, and a lot of other features. A substantial amount of functionality is packed into this $30 a month subscription.

The $100 a month subscription provides a ton of usage. One heavy user is on the $100 plan. Team pricing also exists at the $30 and $100 levels. This joint pricing combines Grok, Cursor, Grokbot, and X.

Growth Claims and Supporting Comments

Elon said Grokbot hits 418 weekly users after just one month. It remains unclear exactly how many people currently use Grokbot each week when separating it from Grok, SpaceX, and related services. Some people say numbers are being hidden behind the merge of the three. Elon then stated he has never seen anything grow as fast as Grokbot. There is no free tier, which means people are deriving real value in excess of what they pay. You get what you pay for.

This was in response to a video clip from Jensen Wang. He said Grokbot usage is skyrocketing. In the video the CEO spent one minute on Elon and Grokbot. He noted that Elon trains and runs his own models. Grokbot demand is climbing fast. He called it super cool.

Jensen was originally a big fan of Clawbot and others that grew really quickly. Those topped out and died very quickly because the real value was not there. Elon’s suggestion is to start small. Give Grokbot one research task that has been put off. Grab the $30 a month plan. A lot of value comes with it. This is not a paid advertisement.

Jensen Wang says he loves working with Elon. He is delighted and loves working with Elon. Elon is using their technology for his own training models and for deploying models. Grokbot is super cool. The combination of these three things is fairly unique. Google does something similar as another example.

One recent interesting exchange involved Tesla’s Roadster launch and flying cars. Elon was posting about it. Sundar Pichai, CEO of Google, popped in and gave compliments to Elon. Elon responded. It is interesting to see these public exchanges. Jensen has also spoken about his collaborations with Elon.

Elon has said Grokbot usage is growing faster than anything they have ever seen. SpaceX Grokbot hits 418,000 weekly users ending September 14th. That figure is from a month ago, so the number has probably increased since then. Some people say the merge of the pricing and related changes has pushed people off, though that remains to be seen.

A Heavy User’s Perspective and Comparison to Metamuse

At Startuphakk, custom software solutions are built for companies. With a decade of executive leadership as a fractional CTO and 25 years in software development, the work includes transforming tech teams and products, including building custom AI solutions.

The team runs Openmonoagent. It is open source and given away free. Anyone can open the code, run it, and install it with one command for a full local setup for AI. A Strata build is being pushed so Openmonoagent can run on top of Strata. On a 5090 with 64 gigs of RAM, the setup is getting around 180 tokens per second on a huge model—not even as small as a 27 billion parameter model. Boundaries are being pushed.

On smaller hardware, work is underway to move from the Quen 3.59 billion model to the Bonsai 2.12 billion and then the Bonsai 2.27B. These run faster on a bigger model using the Bonsai model. The system detects hardware and decides the best model to run, along with the UI and TUI. An agent host mode is also launching. On a development or R&D box, the agent host can be installed to troubleshoot installs, optimize, and support a recursive loop so developers can run faster. About four branches are merging into main with new features.

This context matters for Grokbot usage. Hardware in the R&D lab includes a system running a Bonsai 2 model on a 3060 with 12 gig. Another (a white unit) is running Strata at 180 tokens per second on a huge model such as something in the range of Quen 3.8 Flash Next. This has become a daily driver for a development group on a 5090 with 64 gigs of RAM. A black unit runs a 3090 with 12 gig. Production runs in a private data center, not from the office.

Metamuse, the agent, was tested. The Metamuse models themselves (Wonder and Spark 1.3) are really good and work very well. The agent, however, feels more like a tool for personal or home use—booking flights, hotels, and similar tasks. When given work tasks, it performed poorly in the examples tested. Many work tasks were pushed to it and it did not do well.

When Grokbot was stood up, it integrated into the R&D lab using burner accounts. It is not connected to any production work accounts. A cloud bot is not allowed to touch production systems. Grokbot is managing Openmonoagent installs on the machines in an R&D lab setting and is doing a lot of that work. It is very powerful.

The difference is that Metamuse feels more like a personal tool—stronger than a toy, but oriented that way. Grokbot functions as a work tool and a workhorse. Around 20 different Grokbots are running and doing various things. It is not perfect and is worked with every day. The team is also using it for various tasks. There are many things it does very well. From a business level, the $30 plan is really good.

Retention, Value, and Business Use

Even if there is some churn related to combining Grok, Cursor, and related services, it is better to be loved by a few than liked by many. OpenAI is liked by many, with tons of free users—hundreds of millions, somewhere near a billion free users. The question is how many people want to pay for it. Elon is saying there is no free tier for Grokbot: put your money where your mouth is and pay the $30 a month. Those users are likely to be business and corporate accounts.

In one business, the $100 per month Grokbot, Cursor, and related stack is paid for along with many other services. It is easy to get $100 a month of value out of the business from it. For that business a $100 line item is fairly small. The $100 per month would be kept, and Grokbot is providing that value. This backs up Elon’s statement. Developers are using both the $30 and $100 ranges and getting value out of it. As someone who is careful with business spending, the usage reflects genuine value found in the tool.

SpaceX Quarterly Numbers on AI

SpaceX’s quarterly earnings provide additional context. Their AI revenue was $2.56 billion, up 247% from last year. The quarterly filing breaks it down. $1.6 billion of the jump came from renting out Compute and GrokX subscriptions. Grok and GrokX subscriptions added $258 million. They spent $16 billion on AI buildout in that one quarter and hit 1.4 gigawatts of computing power.

Per the IPO filing, Anthropic pays a reported $15 billion a year for access to Colossus. That $16 billion buildout is already getting paid for in the first year. Elon is renting his data centers to Dario at Anthropic. The same filing shows about 117 million monthly Grok users, but only about 1.9 million paying for SuperGrok. When the numbers from SuperGrok, Cursor, and related services are combined, the picture changes somewhat. As a consumer, more comes out of that combination.

Broader Tool Usage and Advice

Nearly every stack is in use right now except for Codex and OpenAI, which are not used. One team is still on CloudCode because they are in a big release crunch. Many people are on OpenCode. Free options have included Metamuse Spark 1.3 (free up to around a billion tokens) and recent free offerings from DeepSeek and others on OpenCode. Developers are also using OpenMonoAgent for daily coding work, and limits are being pushed with it.

Staying on the cutting edge matters for a business that builds custom software solutions. The $30 a month plan to try Grokbot, which also includes a bunch of Cursor usage and other features, feels like pretty good value.

Startuphakk builds custom software solutions for companies, particularly AI build-outs. Most companies do not have a technology problem; they have a leadership problem. They pay for it in missed deadlines, failed integrations, and AI investments that deliver nothing. Spencer Thomason is a fractional CTO and founder of Startuphakk, with 25 years building real software and a decade in executive leadership at organizations such as GoDaddy, SRP, and Wells Fargo. Bad technology decisions do not just waste money; they can kill companies.

The businesses winning right now are not necessarily the ones chasing the latest AI trends. They are the ones who built on solid engineering and treated AI as infrastructure they own, control, and integrate into software that actually works. The team delivers real custom software development: database architecture, API design, system integration, and scalable infrastructure built the right way. When AI belongs in a solution, it is designed into the architecture and runs in the customer’s environment, not a vendor’s cloud or on someone else’s pricing schedule.

An open-source platform called OpenMonoAgent.ai is a terminal-native AI coding agent running entirely on local LLMs. It has zero API costs, zero telemetry, and full ownership. It is growing because engineers recognize real infrastructure. As a fractional CTO offering, the same standard of leadership is available without full-time executive costs: strategic architecture, hands-on delivery, and clear accountability.

The growth numbers, the lack of a free tier, the pricing changes, and direct usage experience all point to Grokbot delivering measurable value for those who pay for it, particularly in business and development settings. Whether the fastest-growth claim holds over time will depend on continued retention and real-world results after the recent mergers and pricing shifts.

Conclusion

Elon highlighted Grokbot’s rapid growth and the lack of a free tier as signs that users get real value for what they pay. Bloomberg’s 418,000 weekly users and Jensen Wang’s comments on skyrocketing demand support that view, even as the new joint pricing with Cursor and X raises some retention questions.

In practice, the $30 and $100 plans deliver usable capacity and clear business value for teams handling real work, with Grokbot functioning as a workhorse rather than a personal tool. The SpaceX AI revenue figures and related deals further show the commercial context behind the growth. For those evaluating it, starting with the $30 plan on a single deferred task offers a straightforward way to test the claimed value.

The post SpaceX Grokbot Growth: Pricing, User Numbers, and Real Business Value first appeared on STARTUP HAKK.

Anthropic’s Revenue Is Growing, but Compute Costs Are Growing Faster

Spencer Thomason — Thu, 15 Oct 2026 19:16:20 +0000

Anthropic’s quarterly revenue grew 3.6 times this year, rising from $4 billion in the first quarter to $16 billion in the third. That headline number is large. At the same time, the company has locked in far larger commitments for compute that cannot be canceled. By the end of 2025 those bills stood at $54 billion. More recent figures put the total promised compute at more than $518 billion—roughly 9.5 times higher. Revenue expanded 3.6 times while the non-cancelable compute obligations expanded nearly 10 times.

The pattern matches earlier periods. Revenue rose by about $4 billion while costs rose by about $9 billion. For every new dollar of revenue, roughly two dollars went out the door. The gap between money coming in and money committed to go out has widened, not narrowed.

Revenue Numbers and the Difference Between Headlines and Reported Quarters

Public discussion earlier pointed to a $100 billion yearly revenue pace. When Anthropic later shared its actual third-quarter figures with investors, the number was $16 billion, up from $11 billion in the second quarter. One analysis of that quarterly total translated it to roughly $5.8 billion for September alone. Multiplying that figure by twelve produces an annualized pace near $70 billion rather than $100 billion. August and September each contributed only about $2–3 billion of additional annualized revenue—the weakest months of the year under that reading.

The sequence of quarterly figures given was approximately $4.7 billion in the first quarter, $11 billion in the second, and $16 billion in the third. That represents substantial absolute growth, yet the rate of increase has slowed: the second quarter showed roughly 146 percent growth over the first, while the third quarter showed about 45 percent growth over the second. A valuation discussion that has referenced figures above $2 trillion would require continued very high growth rates. The most recent quarterly comparison does not show that pace.

What “Adjusted” Profitability Means in Practice

Anthropic told investors it was profitable on an adjusted basis in the three months to September, for the second quarter in a row. The same presentation referenced an adjusted gross margin of 85 percent, an adjusted operating margin of 11 percent, positive free cash flow after adjustments, and positive adjusted operating profits. The word “adjusted” appears repeatedly in these statements.

Those adjustments remove items such as training costs. Training is a core operating expense of building and improving the models. Removing it, along with other non-cash or run-rate adjustments, produces a different picture from the ordinary definition of profit as revenue minus costs. Under that ordinary definition the company is described as nowhere near profitable and still more than twice the level of costs relative to the revenue figures being discussed.

Public companies must eventually file audited accounts that disclose the full picture. Marketing materials and selective investor updates can emphasize adjusted metrics; a formal S-1 registration statement cannot. The repeated reliance on adjusted figures, shared without full context, is noted as a concerning signal about the underlying economics.

Dependence on Amazon and Google

Nearly half of Anthropic’s sales last year—47 percent—were routed through cloud partners Amazon and Google. Those two companies are also major investors in Anthropic, suppliers of the compute capacity Anthropic has committed to buy, and competitors in the AI market. Anthropic paid roughly $351 million back to the platforms in distribution fees.

Sales flowing through the two partners rose from 11 percent to 32 percent during 2024 and reached nearly half of the total the following year. About 60 percent of cash collections were funneled through these third parties. The relationship is circular: Amazon and Google have invested tens of billions of dollars in Anthropic while Anthropic has signed enormous long-term commitments to purchase compute from them. If either partner were to restrict access, the effect on Anthropic’s revenue and operations would be severe. Anthropic itself has previously restricted access for other users.

Distillation, Growth Curves, and Later Projections

A portion of the earlier rapid growth coincided with heavy distillation activity—customers using Anthropic’s models to help train other systems while still paying Anthropic. When Anthropic moved to limit that activity, including restrictions involving China and other parties that had been significant payers, the revenue trajectory flattened. Growth that had appeared explosive during the distillation period moderated sharply after the restrictions. One reading of the data showed essentially flat annualized revenue from July onward once those effects were removed.

Forward-looking arithmetic offered in discussion of the figures projects continued pressure. One set of rough calculations for year-end 2026 pointed to an exit annualized revenue figure in the mid-80-billion range alongside negative operating income on the order of $25 billion. Extending a similar trajectory to the end of 2027 produced an annualized revenue near $138 billion but operating income negative by roughly $60 billion. The direction of the operating result under those assumptions is worse, not better.

Separate reporting referenced a net loss of $42 billion for 2025 alongside plans to spend $518 billion on cloud compute and infrastructure in coming years. Revenue had grown twelvefold in 2025 to nearly $4.6 billion even while the company recorded more than $8 billion in operating losses. These figures sit alongside the much larger compute commitments already described.

Why the Gap Matters for Valuation and Sustainability

The combination of slowing percentage growth, large non-cancelable compute obligations, and profitability that appears only after substantial adjustments raises questions about the path to the multi-trillion-dollar valuations that have been discussed. Once full audited financials become public, the difference between adjusted presentations and the ordinary revenue-minus-cost calculation will be visible. One view is that the company may not reach a public offering under the level of scrutiny required, or that any public valuation near the highest figures circulated would not hold. Another possibility raised is that the business could ultimately be acquired for its components rather than continue as an independent entity at the scale currently implied by the adjusted metrics.

The core arithmetic remains straightforward: revenue has grown quickly in absolute terms, yet the commitments to spend on compute have grown faster. Adjusted profitability excludes major costs that are inherent to operating the company. The result is a picture in which money continues to leave faster than it arrives.

Conclusion

Anthropic’s revenue rose from about $4 billion to $16 billion, yet non-cancelable compute commitments expanded far faster to more than $518 billion. Profitability appears only after adjustments that remove core costs such as training. Under ordinary revenue-minus-cost accounting, money continues to leave faster than it arrives, raising questions about whether the current trajectory can support the valuations under discussion.

The post Anthropic’s Revenue Is Growing, but Compute Costs Are Growing Faster first appeared on STARTUP HAKK.

Anthropic Cuts Claude Reasoning Access by 99% to Stop AI Copycats

Spencer Thomason — Wed, 14 Oct 2026 18:59:33 +0000

Anthropic has publicly stated that a substantial share of its paid traffic came from other AI labs attempting to copy Claude. The company turned away trillions of tokens worth of that traffic. On Monday it posted that a new defense reduced the amount of Claude’s thinking copycat labs can pull by 90 to 99 percent. It banned tens of thousands of accounts, many of them within two hours of starting, and named a number of labs coming out of China. One network operated more than 20,000 fake accounts at the same time using stolen payment details and fake identities. Anthropic has acknowledged that the action meant turning away meaningful revenue.

How Distillation Works and Why Anthropic Hid Claude’s Thinking

Distillation means taking a large, expensive model such as Claude, sending it millions of questions, and saving every answer along with every bit of its reasoning. A smaller model is then trained on that collected material. The approach functions as a shortcut that skips years of work and billions in compute by copying someone else’s homework.

Anthropic’s own tests illustrate the strength of the shortcut. A model trained on Claude’s full reasoning reached 80 percent on a popular coding benchmark. When the same model received only a summary of the reasoning, the score fell to 71 percent. That difference is why Anthropic began hiding Claude’s full thinking process. Paying customers do not see it either.

Scale of the Activity and Named Labs

Anthropic described the activity as covert industrial-scale campaigns. One example cited is Xifu, which distilled cyber capabilities of leading U.S. models during development of GLM 5.3. The same company tried to distill Claude Fable 5; after Anthropic’s cyber safeguard degraded those attacks, it switched to an older Claude model. Other references point to Zaipu and Moonshot.

The new measures control where reasoning can go, block and patch extraction attempts, and make repeated attempts difficult. A graph shared by Anthropic showed average attempts dropping sharply once the safeguard was fully applied, with some distillation attacks falling to zero. The company expects the contest to remain a cat-and-mouse game that will require continued iteration of defenses. Distillation is framed as a problem for every developer platform. Despite the tradeoffs, Anthropic said the measures are clearly worth the cost for reasons of national security and ensuring power.

Revenue Questions and the Timing of Charges

Because many of the accounts used stolen cards and fake identities, part of the usage growth that had been celebrated appears to have come from copycats. Anthropic is seeking to go public this year, so the reduction in traffic raises questions about the numbers that supported earlier projections. Since September 24, Anthropic has charged customers for certain requests that are blocked before Claude even responds, meaning payment can occur for a refusal. The new classifier carries a false-positive rate of 0.01 percent and is described as roughly as robust as the company’s safeguards against bioweapons assistance.

One AI skeptic asked directly how much of Anthropic’s revenue had been coming from distillation. The accounts in question were not free sign-ups; they were API customers routing substantial volume through the service. Turning that volume away is expected to affect the bottom line. Average weights on some days dropped from 8,000 to 135 times, and the remaining leak became too small to measure.

Effects on Other Models and Industry Projections

The reduction in accessible reasoning raises questions about how models such as GLM 5.3 and those associated with Moonshot and Kimmy will continue to perform if distillation becomes far harder. Earlier market reactions to inexpensive Chinese models had been linked to the use of distillation accounts. Anthropic stated that it had exposed foreign labs conducting industrial-scale extraction of Claude’s reasoning through hundreds of thousands of fake accounts.

Broader projections for AI spending have been cited as context. Americans may need to spend around $3.5 trillion a year on AI services by 2032—five years from now—to justify current data-center build-outs. That figure equals 8.8 percent of GDP. Present spending is described as not even at 1 percent and possibly not half a percent. The build-out itself is projected to cost $10 trillion from 2025 to 2032, a larger share of the economy than the railroad, highway, or telecom booms. Demand for the underlying hardware continues to climb, yet the requirement for spending at the scale needed to sustain current pricing ratios is viewed as unlikely to materialize fully.

Open-weight models have begun to grow. The prediction offered is that Anthropic’s revenue will table off rather than continue the doubling and 10x growth seen over the last 18 months. Whether the company can complete an IPO on the basis of earlier growth assumptions remains an open question.

Local Infrastructure as an Alternative Approach

Startuphakk builds custom software solutions and has used Open Mono Agent, an open-source repository designed for local AI. The preference is to avoid sending data into the cloud. Local setups eliminate a usage meter and keep infrastructure under the owner’s control—the same principle applied for 30 years. Current testing fits models into 12 gigabytes, allowing use of graphics cards consistently available for $300 to $400 rather than the $2,000 to $3,000 range. The platform emphasizes local-first operation, unlimited tokens, sandbox by default, and open source.

Most companies are described as facing failed integrations and AI investments that deliver nothing rather than pure technology problems. Starterpack’s work focuses on database architecture, API design, system integration, and scalable infrastructure built the way it has been done for more than 25 years. When AI is included, it is designed into the architecture running in the client’s environment rather than as a wrapper around someone else’s API or pricing schedule. The openmonoagent.ai platform is a terminal-native coding agent that runs entirely on local LLMs with zero API costs, zero telemetry, and full ownership.

The reduction in accessible Claude reasoning and the associated revenue impact illustrate how quickly the rules can change when a business depends on a model that others can copy. Integrations can break, and the meter can continue to run even when no answer is returned. The episode leaves open how both the targeted labs and Anthropic’s own growth trajectory will adjust.

Conclusion

Anthropic’s new defense cut the amount of Claude’s reasoning that distillation campaigns can harvest by 90 to 99 percent and turned away trillions of tokens of paid traffic. Tens of thousands of accounts were banned, often within hours. The company accepted the loss of meaningful revenue and has begun charging for certain blocked requests.

The move raises questions about both the future strength of models such as GLM 5.3 and Anthropic’s own revenue path ahead of a potential IPO. Rules can change quickly when a business depends on a model others can copy, and some responses focus on local infrastructure that removes the meter and keeps AI under the owner’s control.

The post Anthropic Cuts Claude Reasoning Access by 99% to Stop AI Copycats first appeared on STARTUP HAKK.

Microsoft Brings Local AI Agents to Windows PCs with NVIDIA’s Help

Spencer Thomason — Tue, 13 Oct 2026 18:17:28 +0000

Microsoft wants AI agents living on your own PC. At a Wednesday morning event in San Francisco, Satya opened by describing the goal of bringing unmetered intelligence to every desk and every home. Jensen Wang stood right next to him on stage. The plain-English version of the message is that Microsoft aims to put agents on your computer so they can do work for you without a cloud bill ticking every time the system thinks.

Microsoft’s Own Coding Model for Local Use

The coding brain chosen for the desk is not from OpenAI. It is Microsoft’s own model, MIA Code 1.1 Flash (also referred to as Maya Code Flash). This is a 137-billion-parameter coding model with a 256K context window. It has been squeezed down almost 80 percent so it can fit onto a laptop and is now optimized to run on your PC.

GitHub Copilot can hand off a lot of the work to the local model through hybrid intelligence. Hybrid means the system is partial cloud and partial local. With code and Copilot, the idea is that you can essentially build any software you need on your desktop without any cloud token spend. At the same time, the hybrid approach keeps Microsoft’s involvement in place.

Performance Claims and the Memory Reality

On stage, Microsoft showed a slide claiming up to 6.2 times faster performance than a MacBook Pro. The tiny print noted that this was based on preliminary testing by NVIDIA. The catch is that the coding models peak out at 75 gigabytes of memory. A Surface with 24 gigs is listed in the $25,000 range in the discussion, while the 128-gig configuration is described as 5,900 (with other references to a high-end Surface Pro around 59,000 and nearly six grand for the top memory option). Microsoft’s own developer page indicates the router moves up to 20 percent of Copilot’s work to local models. In a live demo, the hard prompt still went to GPT-6 Luna in the cloud while small subagents ran on the local machine. The grunt work that burns most tokens is what Microsoft wants running on hardware you buy, using a model Microsoft owns.

NVIDIA’s Partnership and the Hardware

Jensen described how Windows sparked a platform shift that created a new industry for NVIDIA. The company invented programmable shading GPUs for DirectX, which led to CUDA. It then partnered to bring GPU supercomputers to Azure, which helped OpenAI train GPT. That collaboration inspired the effort to reinvent Windows for the age of personal agents. Four years and thousands of engineers went into the work between the two companies.

The new devices include the Surface Laptop Ultra, a 15-inch machine powered by the NVIDIA RTX Spark Superchip with up to 128 gigs of unified memory. RTX Spark puts NVIDIA’s Grace ARM CPU and Blackwell RTX GPU on a single Superchip. The top configuration has 20 CPU cores and 6,000 CUDA cores. A lower 18-core model drops to 51 GPU cores and is associated with 24- or 32-gig options. At 32 gigs with the higher core count, the setup can run a model like QUEN 3.8 at 27 billion parameters pretty well. The 24-gig version will not get you there, but 32 gigs plus roughly 5,000 cores supports a solid 27B run. Because the CPU and GPU share unified memory, a large AI model can use far more than the VRAM it would typically sit on. Plenty of disclaimers appear in the fine print.

Security, Sandboxes, and Who Controls the Agent

An agent on your PC runs with your access, your files, your emails, and your logins. Microsoft made its own sandbox—Microsoft execution containers—generally available on Windows 11. Codex from OpenAI already supports it, and Cloud Code is on the way. When both Sam and Dario have to plug into Microsoft’s sandbox to run on Windows, the question of who the landlord is becomes relevant. Security is described as foundational, bringing Windows and Agent 365 together so agents can stay secure. The official positioning calls Windows the most secure platform for agents, even though Windows has not always held that reputation.

The Hedging and the Double-Dip

Microsoft is hedging off OpenAI one box at a time. The company that spent years tied closely to Sam Altman is now putting its own model on the machine for a portion of the work. Sam keeps the hard questions while Microsoft goes after the volume. Jensen gets paid either way: the same chips that helped train GPT in the cloud now sit inside the laptop. NVIDIA sells the shovels on both sides of the fight. Over half a million people have already seen related discussion of the announcement.

Local AI as Infrastructure You Own

The broader thesis is that local models, a harness around an agent, and a sandbox on hardware you own put control in your hands. Microsoft’s announcements highlight thousands of actions from the taskbar on one laptop, with a bigger desktop available for heavier grunt work. Hybrid intelligence brings frontier capability to Windows locally and supercharges Copilots. At the same time, the 20 percent local figure and the continued cloud path for harder tasks show the limits of the “unmetered” claim.

People who have been building on local models for well over a year see the value in machines that keep data on the device rather than continuously pumping it to the cloud. Current high-end setups already include RTX cards such as a 5090 paired with a 3060 or a 3090 that remains relevant years later. The new unified-memory Windows hardware raises the possibility of running something like a 27B model comfortably and pairing it with local-first tools. One open-source approach already in use is OpenMonoAgent.ai: a terminal-native AI coding agent that runs entirely on local LLMs, is sandboxed by default, Docker-native, mounts the project with permission gates on destructive operations, and offers zero API costs and zero telemetry. It has been positioned as local-first and unlimited-token from the start.

Conclusion

Microsoft’s event with NVIDIA frames a new chapter for Windows in which every PC becomes a place where agents can work securely on your behalf. The company is pairing its own compressed coding model, hybrid handoff of Copilot work, and new RTX Spark hardware with its execution containers. NVIDIA’s chips sit on both sides of the cloud-training and local-inference equation. The practical outcome is more work staying on hardware the user buys, even while hard prompts still route to the cloud and memory requirements remain high. Whether that leads to broader local ownership of AI infrastructure or simply another set of Microsoft-controlled devices is the question the announcements leave open.

The post Microsoft Brings Local AI Agents to Windows PCs with NVIDIA’s Help first appeared on STARTUP HAKK.