AI agents don't see ads. They just bypass them entirely. The internet is built on advertising revenue from human eyeballs. Agents don't have eyeballs. Now a Coinbase engineer has a pitch… if a human visits, show them an ad. If an agent visits, charge them five cents. Five cents though. Do the math. One task might mean a hundred pages scraped. That's five dollars per task at machine scale. The economics need to match how agents actually operate. “Agents really are the browser of the future.”

Nvidia just invested $500 million in Corning, a glass and fiber company. Let that sit for a second. Nvidia, the company that makes the chips everyone fights over, is spending half a billion on glass and fiber. Not more GPUs. Glass and fiber. The bottleneck for AI is not just the compute layer. The physical infrastructure is equally constrained. Fiber optics, glass substrates, networking hardware, these are the pipelines that move data between chips and data centers. When the pipelines are full, the chips sit idle. Corning makes the specialized glass used in semiconductor manufacturing and the fiber optics that connect everything. Nvidia locking in $500 million of supply tells you the physical infrastructure is as tight as the silicon. We have been tracking AI infrastructure all day. Panthalassa building ocean-based power. Goldman Sachs noting compute infrastructure is driving inflation. Nvidia now spending on the supply chain layer, not the chips, the glass. The AI buildout is hitting physical constraints at every level. Energy. Compute. Networking. Infrastructure. This is what supply chain bottlenecks look like when they become strategic.

Goldman Sachs is connecting AI to inflation. Three vectors: Energy consumption. Compute infrastructure. Labor market premiums. The mainstream story is that AI reduces costs, efficiency gains, automation replacing expensive labor, lower prices for everyone. Goldman Sachs is pushing back on that in the short term. The buildout phase is inflationary. You have to build the power grid, the data centers, the GPU clusters, and pay the specialists before you get the efficiency. Energy costs are not theoretical. Data centers are bidding against manufacturing and residential consumers for electricity. Compute hardware, GPUs, cooling systems, networking, is expensive and supply-constrained. AI specialists command salaries that reflect the scarcity of the skill. Three cost vectors, all pushing prices higher while the efficiency gains are still being built. This is not a permanent argument against AI. It is the short-term reality of infrastructure buildout. The costs come first. The savings come later. Goldman Sachs is not a natural AI skeptic. They are reading the inflation data and following the money. Right now, the money is flowing into buildout, not into savings.

Peter Thiel just backed $140 million into something called Panthalassa, floating ocean nodes that generate wave power and run AI servers in the same structure. Let that sink in. Not wave power feeding a grid that feeds a data center. Wave power generated right where the AI servers sit. No transmission lines. No grid. No transmission losses. Portland-based startup. Steel nodes the length of a football pitch, floating in open ocean. AI inference running on wave energy. Cooled by seawater. Deploying pilot nodes in the northern Pacific. Why this matters beyond the novelty: AI's energy appetite is forcing infrastructure innovation that does not look like traditional infrastructure. Land-based power generation, transmission lines, substations, data centers that supply chain has limits. Permitting, grid capacity, land availability. The ocean sidesteps all of it. Wave power has been a promising technology for decades. It never scaled because there was no customer big enough to absorb the energy at the cost point needed. Now there is. AI compute demand is so large that ocean-based power, expensive, mechanical, maintenance-intensive, finally has a buyer. The ocean provides power and cooling. The data center is at the power source. No grid dependency. This is what AI's infrastructure phase looks like when it gets serious.

A Miami startup just announced SubQ, claiming the first fully subquadratic sparse-attention architecture in a frontier LLM. 12 million token context window. 52 times faster than FlashAttention at 1 million tokens. 50 times cheaper. Potential 1,000x efficiency gain. Here is why the numbers matter. Current LLMs use quadratic attention, double the input, quadruple the compute. That's the fundamental bottleneck that makes long contexts expensive. Subquadratic means compute grows linearly. Double the input, double the compute. Same cost regardless of length. If it works, you could process an entire codebase, a company's full documentation, hundreds of research papers, all at the same cost per token. That is a different category of capability. The important caveat.. researchers are asking for independent verification. Bold claims from a new company. No third-party benchmarks yet. SubQ could be a genuine leap or it could be overpromised. We have been saying this all morning… AI is accelerating faster than the public narrative reflects. A 12 million token context is not incremental improvement. It is a different category. Whether SubQ delivers is a separate question. The direction is clear.

Jack Clark just put a number on the thing nobody wants to talk about. Co-founder of Anthropic before he left. Now he's estimating 60% probability of recursive self-improvement in AI by 2028. That means AI systems that can improve their own capabilities without human input. AI that designs the next generation of AI. Not next decade. Next two years. The distinction from where we are now cannot be overstated. Current AI, even the most capable models, still depend on human engineers to improve them. We write the code, we design the architecture, we train the next version. The capability ceiling is set by us. Recursive self-improvement is different. The system gets better at improving itself. The rate of progress becomes self-sustaining. The human in the loop becomes optional. Eventually, maybe irrelevant. Clark is not a pessimist. He helped build one of the leading AI labs. If he's putting 60% odds on this in two years, the people inside the labs know something. This is not the AI that answers questions. This is the AI that answers the question of what comes after questions. Two years. 60% odds.

Apple just agreed to pay $250 million for falsely claiming AI-powered Siri was available when it was not. The case covers 36 million iPhones sold between June 2024 and March 2025 based on AI capabilities that did not exist, did not arrive, and according to the lawsuit, will not exist for two or more years. Apple ran a major advertising campaign promoting these features. The BBB's National Advertising Division concluded Apple falsely suggested the new Siri was available now. The enhanced Siri was the feature iPhone buyers most anticipated. Morgan Stanley research showed it. Apple marketed it. People bought on it. $250 million sounds like a lot until you realize Apple makes $391 billion in revenue annually. This is roughly what they earn in 6 hours. The real question is whether this settlement becomes a template for AI advertising accountability, or just the cost of doing business for companies that oversell AI. The AI hype machine is running into real consequences. Fines. Lawsuits. People in jail from AI errors. Hallucinations destroying reputations. Promises that never shipped. Accountability for AI claims is coming. Slowly. But coming.

Coinbase is cutting 14% of its workforce. The CEO says it's becoming an AI-first company. That's roughly 660 jobs from a staff of 4,700, going in an email to employees today. The same company that launched an AI agent marketplace in April, where hundreds of thousands of agents transact hundreds of millions, is now rebuilding itself around the technology it was selling. This is the pattern now. Meta cut 8,000 jobs while doubling AI spend to $135 billion. Shopify achieved 100% AI tool adoption across its entire workforce. Snap cut 16% of staff for AI. The common thread isn't just that AI is changing work, it's that companies are choosing AI over human workers at scale. What's different here is the company doing the choosing. Coinbase sits at the intersection of crypto and AI infrastructure. It built the machine-to-machine payment layer that AI agents use to transact. It integrated AI deeply into its operations. And now the same technology that powered its growth is reducing its headcount. That irony is the story. The tools being built are replacing the people who build them.

Retail traders are putting AI agents in charge of their portfolios. Bloomberg reported on May 1 that ordinary traders, not hedge fund engineers, are now deploying autonomous agents to research markets, debate trade ideas, and execute strategies across equities, crypto, and prediction markets simultaneously. One trader's bot refused to chase Nvidia momentum post-earnings. The human would have piled in. The bot held because its parameters flagged the risk profile and stayed within them. That is not the AI being smarter. That is automation enforcing discipline human psychology undermines at the worst moments. The architecture is straightforward: a language model does research, a rules engine makes decisions, an execution layer connects to brokerage APIs. Some traders run multi-agent debate setups… one AI argues for a position, another argues against it, the human reviews before acting. What makes this significant is the feedback loop. Bad enterprise decisions waste time. Bad trades lose money immediately. That clarity makes retail finance one of the most useful proving grounds for autonomous AI systems available, and the lessons being learned now will shape agent design well beyond the trading desk. The democratization of finance is not a tagline. Retail traders with accounts on Robinhood, Coinbase, and Polymarket can now run a single agent across all three, something that required institutional infrastructure five years ago. The results are uneven. The direction is not.

American Express just made AI shopping official. Their new ACE kit, Agentic Commerce Experiences, lets AI agents shop and pay on behalf of users. Not a concept. A working system with intent contracts, single-use tokens, and a commitment to cover erroneous transactions made by registered AI agents on their network. The liability question has been the gap. When an AI agent buys the wrong thing or gets manipulated, who holds the loss? Amex is answering… we do. That precedent changes what autonomous commerce can look like. We covered Circle's USDC nanopayments and the agentic commerce infrastructure building out. The payment rails are being laid. Amex is now putting the liability scaffolding in place. AI agents that can transact. AI agents that have infrastructure to do it securely. AI agents with someone standing behind them when it goes wrong. The agentic economy is not coming. It is arriving in parts.

Haun Ventures just raised one billion dollars for the crypto-AI convergence. We have been saying it for weeks now. The infrastructure layer and the intelligence layer are not separate trades. They are the same bet. Bitcoin's payment rails. AI agents spending money autonomously. DeFi protocols running without intermediaries. Financial infrastructure and cognitive infrastructure becoming one system. Haun's thesis is exactly what we have been building toward… the point where Bitcoin provides the financial plumbing and AI provides the autonomous decision-making that makes the whole thing function. Payment rails plus agents. Sound money plus intelligent execution. When one billion dollars of institutional capital arrives at the intersection you have been teaching about, the thesis stops being speculative. It becomes the mainstream position. The convergence is not coming. It is already here.

The Trump administration is considering pre-release vetting for AI models before public launch. The same administration that killed Biden's AI safeguards is now drafting an executive order for government review before new models ship. Formal oversight. Approval gates. The works. But here is the question nobody is asking… how does vetting new frontier models control the AI already everywhere? Open-source models are already running on millions of machines. Thousands more sit on Hugging Face. International providers operate outside US jurisdiction. AI is embedded in products sold and deployed. Vetting only covers new models from major labs willing to play. It creates friction for the top of the market while the base already runs without oversight. The proposal is a narrative about control. Not a mechanism for achieving it. The diffusion happened. The AI is already out.

Microsoft just launched Agent 365 to manage shadow AI. Shadow AI is when employees deploy AI tools without IT oversight, and it is growing fast. Microsoft's framing is direct.. this is the new shadow IT. Companies deploying AI without governance is an enterprise risk. But here is why shadow AI is different from shadow IT. Shadow IT meant unauthorized software storing data. Inconvenient. Fixable. Shadow AI means unauthorized AI reading data, drafting contracts, generating code, writing customer responses and those outputs getting acted on without anyone checking. The velocity is different. The blast radius is different. A shadow AI tool that an employee uses to summarize a client database is not the same as a shadow spreadsheet. The AI can act. It can infer. It can expose. Microsoft is not wrong to be worried. Agent 365 is their answer. But a product that manages shadow AI still accepts that shadow AI exists. The real question is why employees feel they need to go around IT in the first place. That gap.. between what is approved and what people actually use, is where the risk lives. And it is not shrinking.

The AI hallucination problem is escalating. We started with lawyers fined for citing fake cases. Journalists suspended for publishing fabricated quotes. Now we're seeing real-world harm with real consequences. Ashley MacIsaac, a Juno award-winning Canadian fiddler, is suing Google for $1.5 million after AI Overview falsely claimed he was convicted of sexual assault, internet luring of a child, and assault causing bodily harm. A First Nation cancelled his concert based on the misinformation. He says he feared for his safety performing. The legal argument is critical.. Google should not have lesser liability because the defamatory statements were published by software that Google created and controls. But there's a worse case already on record. Angela Lipps, a Tennessee grandmother, spent six months in jail after facial recognition AI incorrectly identified her as a bank fraud suspect. She was arrested at gunpoint while babysitting four children. She'd never been to North Dakota where the alleged crime occurred. She spent four months without bail awaiting extradition. She lost her home, her car, her dog. Fabricated quotes can be corrected. A reputation damaged can be rebuilt. Six months in prison cannot be undone. AI errors used to feel like technical problems. Software bugs. Edge cases. Now they're putting people in jail and destroying lives. The legal accountability question isn't abstract anymore. It's urgent.

Someone just prompt injected an AI using morse code. The attack worked. Over $200,000 in tokens moved before the market recovered. We've discussed prompt injection before. The uncomfortable truth is it's not solvable through detection. You can filter keywords, add guardrails, fine-tune, but the creative attack surface keeps expanding. Morse code was just used as a bypass. The only real solution is architectural isolation. Capability-based permissions. Narrow scoping. Systems that can't execute unauthorized actions even if compromised. As AI gets integrated into financial infrastructure, this matters more than ever. The morse code attack was a proof of concept. Next time it won't be DRB tokens. It will be whatever the next integration point is. The security model has to assume the AI can be compromised. Build from there.

$10 trillion in monthly stablecoin volume. That's happening right now, and it barely registers in mainstream crypto coverage. The stablecoin infrastructure is quietly becoming the largest payment network most people have never heard of. Trillions moving through USDT, USDC, and the rest, settling transactions globally, every single month. Meanwhile Bitcoin just cleared $80,000 while the AI infrastructure buildout accelerates and Morgan Stanley quietly positions for Bitcoin custody integration. The fundamentals aren't waiting for the headlines.

The AI narrative has always been about open innovation and democratized intelligence. But the actual driver right now is defense contracts. Oracle just joined the list of Big Tech companies aligning with Pentagon national security efforts. The biggest AI development isn't happening in research labs or startups. It's happening under government partnership. When the Pentagon is your biggest customer, your incentives shift. Your priorities shift. The "open and empowering" story faces real pressure. Sovereign AI capabilities are becoming a strategic necessity for nations that don't want to depend on defense-adjacent providers. The AI landscape is consolidating under state influence. What comes next won't look like the original vision.

Nvidia just moved 90% of its production costs to Asia. One year ago that number was 65%. The geopolitical shift is structural, not temporary, and it's happening in real time. Companies in South Korea, Taiwan, and China are deepening ties with Nvidia: SK Hynix, Samsung, LG Electronics. All seeing rallies as the supply chain reshapes itself. The US-China tech war is pushing Nvidia to diversify. Export controls are forcing the rebuild. This isn't a temporary arrangement, it's the new architecture of AI infrastructure. Prediction markets are pricing Nvidia at 68.5% odds to become the world's largest company by June 2026. That's not speculation. That's markets pricing the supply chain reality.

Apple is using private iPhone data to train its AI. The company that built its brand on privacy is now using your messages, photos, and usage patterns to train Apple Intelligence. Today we reported on CNN, NBC, and USA Today fighting for their content rights. Now it's your data too. The AI training data debate has two fronts: creators who made the content, and users who generated the usage patterns. Both are waking up to the same reality, their data trained systems they didn't consent to. The privacy questions are valid. The regulatory friction is coming. This is the consent gap at scale. Not just copyrighted articles. Not just browsing data. Your private messages.

Apple is using private iPhone data to train its AI. The company that built its brand on privacy is now using your messages, photos, and usage patterns to train Apple Intelligence. Today we reported on CNN, NBC, and USA Today fighting for their content rights. Now it's your data too. The AI training data debate has two fronts: creators who made the content, and users who generated the usage patterns. Both are waking up to the same reality, their data trained systems they didn't consent to. The privacy questions are valid. The regulatory friction is coming. This is the consent gap at scale. Not just copyrighted articles. Not just browsing data. Your private messages.