ANTHROPIC GREW 80 TIMES FASTER THAN IT PLANNED AND HAD TO RENT MUSK’S DATA CENTER TO KEEP UP
So here is a number that does not make any sense until you sit with it for a minute. Anthropic planned for ten times growth this year. They built spreadsheets, modeled scenarios, hired accordingly, bought compute, all of it. And then reality showed up and did eighty times. Not ten. Eighty. Dario Amodei himself stood on stage at the company’s Code with Claude developer conference and called it “just crazy” and “too hard to handle.” For a guy who usually talks like he is testifying before Congress, that is a significant departure from his normal register.
The numbers are genuinely hard to process. Anthropic’s revenue run rate hit $30 billion in April 2026. Last December it was $9 billion. In January 2024 it was $87 million. To put that in perspective, Salesforce took twenty years to reach $30 billion in annual revenue. Anthropic got there in under three years from basically nothing. The engine driving all of this is Claude Code, the company’s agentic AI coding tool, which hit $1 billion in annualized revenue within six months of launching. Developers who use it are now averaging twenty hours a week with the tool. Anthropic’s own engineers now write the majority of their company code using Claude Code. The product that pays the bills is also the product they use to build the product. That feedback loop is almost impossible for competitors to replicate without something comparable to close the gap.
The enterprise numbers are equally striking. Anthropic now has over 1,000 customers spending more than a million dollars a year on Claude services, a figure that has doubled since February. Uber, Netflix, and a wave of other major companies are in that group. Amodei’s framing at the conference was blunt: software engineers are the fastest adopters of new technology, and what is happening in AI coding right now is a preview of what is about to happen across the entire economy.
But here is where the story gets genuinely strange. Growing at eighty times your plan is a great problem to have, until you realize you are completely out of computers. Anthropic has been signing deals everywhere to get more GPUs. Amazon, Google, Microsoft, you name it. Most of that capacity will not come online until late 2026 or 2027. They needed compute right now. And so, in a twist that nobody had on their bingo card, Anthropic signed a deal with Elon Musk’s xAI to rent the entire Colossus 1 data center in Memphis, Tennessee. All of it. Over 220,000 Nvidia GPUs. The monthly tab: $1.25 billion. Per month. Through May 2029. That is potentially over $40 billion going directly to Musk’s operation.
Now, Musk spent the last year calling Anthropic everything from “doomed” to actively hating Western civilization. And yet, once Anthropic’s team showed up and shook hands, he changed his tune within a week. He was posting on X that everyone he met at Anthropic was “highly competent” and that “no one set off my evil detector.” So the grudge match is officially tabled, at least while the invoices keep arriving.
The practical reality is that xAI built Colossus 1 for Grok, and Grok’s user numbers cratered after some very unfortunate AI-generated images earlier this year. That left an enormous, expensive data center mostly idle. Anthropic needed compute immediately. Musk needed revenue. The enemy of my enemy becomes my $15 billion per year compute landlord. The AI industry is nothing if not creative about solving its own crises.
Source: VentureBeat | TechCrunch
JENSEN HUANG JUST ANNOUNCED A $200 BILLION MARKET NOBODY KNEW EXISTED AND SAYS NVIDIA IS ALREADY SELLING INTO IT
You have to appreciate the consistency of Jensen Huang. The man sells chips, he makes money selling chips, and he is absolutely unafraid to tell you exactly why you need more of them. At Nvidia’s most recent earnings call, after posting $81.6 billion in quarterly revenue, Huang announced that he has found a brand new $200 billion market that Nvidia has never addressed before. The product that supposedly unlocks it is Vera, Nvidia’s new CPU, introduced back in March. Wait, a CPU? From Nvidia? That is like McDonald’s announcing a salad that is actually good and then casually mentioning it is already doing $20 billion in sales this year.
Here is the logic, and it holds together better than you might expect. When an AI model thinks, generating complex reasoning and output, that work runs on GPUs. That is Nvidia’s whole universe. But AI agents do not just think. They act. They browse the web, run code, call APIs, move files around, send emails. That execution work runs on CPUs, which until now has been Intel and AMD territory. Huang’s argument is that as the world builds billions of AI agents, each agent will essentially need its own computing environment, something like what every person has with a laptop or a phone. Vera is designed for exactly that. It processes tokens as fast as possible, which is what agents specifically need, rather than what traditional CPUs do, which is run many different applications at the same time.
The market has not been entirely convinced that Nvidia can compete on CPUs. Amazon Web Services just signed a massive contract with Meta for millions of Amazon’s own homegrown AI chips. Andy Jassy has been publicly confident that AWS can build chips as good as or better than Nvidia. Intel and AMD have decades of CPU expertise and are not sitting still. But Huang’s answer is not a whiteboard argument. He says Nvidia has already sold $20 billion worth of standalone Vera CPUs in 2026, and it is barely May. Every major cloud provider and systems manufacturer is apparently on board.
His long view is that the world will eventually have billions of AI agents the way it now has billions of personal computers, and each of those agents will need a place to run. Vera wants to be that place. If even a fraction of that vision plays out, $200 billion in addressable market is not hype, it is just arithmetic. Wall Street has been anxious about what eventually knocks Nvidia from its perch. Huang’s answer this quarter is: probably nothing, because we just opened a whole new product category before anyone else could. Coming from the guy who has been right about every major chip transition for the past decade, that is harder to dismiss than it sounds.
Source: TechCrunch
SECRETIVE AI STARTUP WITH NO PRODUCTS JUST RAISED $700 MILLION AT A $6 BILLION VALUATION AND IS BARELY TALKING ABOUT WHY
In the current AI funding environment, $700 million is a lot but not a shock. What makes Hark genuinely interesting is not the number. It is the confidence with which the company raised it while telling almost nobody what it is actually building.
Brett Adcock, the founder, also built Figure.AI, the humanoid robot company, and Archer Aviation, an electric aircraft startup. The man has a documented history of raising enormous amounts of capital for things that sound impossible and then making them real. When he launched Hark in late 2025 by putting $100 million of his own money in before taking a single outside dollar, people paid attention. Now a Series A led by Parkway Venture Capital and backed by Nvidia, AMD, Intel Capital, Qualcomm Ventures, Salesforce Ventures, ARK Invest, and several others has valued the company at $6 billion before the public has seen a single product demo, a product photograph, or honestly much of anything at all.
What Hark says it is building is a “universal AI interface,” an agentic AI system that integrates with all your existing products and services, paired with hardware devices built specifically for those models. The design director is Abidur Chowdhury, a former Apple executive who worked on the iPhone Air. He is being careful to the point of comedy about what he will reveal. Asked directly about the product in an interview, he smiled and said that the team’s demos had impressed investors. That is the entire public disclosure. That is all you get.
What he did say is worth taking seriously. His argument is that almost nobody right now is building AI for regular people. Anthropic is focused on coding tools. OpenAI is racing toward an IPO with developer products. Almost every AI hardware startup is building for enterprise or for builders. Chowdhury thinks there is a massive unaddressed market for AI that actually helps normal humans navigate their daily lives. Not developers. Not IT managers. The person who does not know what a prompt is and does not care to learn.
The problem he is quietly pointing at is real. The AI products that exist today are impressive if you know how to use them. For most people they are confusing. ChatGPT is better than Google for certain things if you know which things. Claude is great for writing if you know how to direct it. But nobody has built the thing that your grandmother picks up on day one and genuinely finds useful without reading a tutorial first. That is apparently what Hark is aiming at.
The privacy and ambient awareness questions are genuinely hard and not yet solved. Wearable AI products have not broken through to mass adoption. Adcock and Chowdhury are clearly aware of all this, which is probably why they are not showing the product yet. They want to solve it before they reveal it. First models drop this summer. That is when the mystery starts to get an answer.
Source: TechCrunch
GOOGLE REDESIGNED SEARCH AROUND AI AND THEN IMMEDIATELY BROKE IT ON THE WORD “DISREGARD”
This one is a gift. After Google I/O 2026, where the company announced the biggest overhaul to Search in its 25-year history, rolling out AI summaries front and center and basically burying the ten blue links that built its empire, the internet noticed something quietly hilarious. The word “disregard” now breaks Google Search.
Not breaks in a dramatic, error-code, crash-the-browser way. It breaks in a quiet, embarrassing, and deeply revealing way. When you type “disregard” into the new Google Search, the AI summary block generates a massive wall of empty white space and shoves the actual results so far down the page that most users on a standard screen never see them. You type a completely normal English word into the world’s most-used search engine and get back a polite void.
The leading theory is straightforward once you understand how adversarial prompting works. “Disregard” is one of the most common trigger words in AI jailbreaks and prompt injection attacks. The phrase “disregard previous instructions” is essentially AI graffiti, plastered across corners of the web by people testing the limits of language models and trying to get them to ignore their guidelines. So when Google’s AI model encounters the word in a search query, it apparently decides the safest move is to generate nothing as a defensive measure. The page then just sits there with a peaceful blankness where your answer should be.
This is funny on several levels. First, “disregard” is a completely ordinary English word. People look it up constantly. For definitions, legal usage, synonyms, context. The dictionary entry for it presumably still exists somewhere below the white space, waiting patiently. Second, it is a live demonstration of the core tension in making a large language model the primary interface for a search engine. The model has to handle both innocent queries and adversarial ones simultaneously, in real time, without being able to reliably tell the difference between someone wanting a definition and someone trying to attack the system. Third, it shows how differently a model-powered search fails compared to a link-based one. The old Google was essentially invulnerable to this kind of confusion. You typed a word and got links. Simple, mechanical, reliable. The new Google has a brain, and brains can be tricked into silence.
Google will fix this specific case quickly. A single trigger word is a tractable problem to patch. But the incident is a preview of the category of challenge that AI-powered search is going to face permanently. Prompt injection is a hard problem that nobody has fully solved, and every AI company is dealing with versions of it. Most of them are not, however, sitting between a billion daily users and the basic act of looking up what a word means. Google made a massive bet that AI summaries would be better than links for most queries. For the word “disregard,” on day one of the new era, that bet produced nothing at all. The word is still technically searchable if you scroll far enough. But eliminating the need to scroll is exactly what the entire redesign was built to accomplish. One English word and the whole premise was briefly, embarrassingly on display.
Source: TechCrunch