The Frontier

Brian Armstrong states that traditional payment rails cannot support agentic commerce because 76% of AI agent transactions are under $0.30, while standard debit or credit card transactions impose a $0.30 flat fee minimum.

Brian Armstrong explains that Coinbase is banking AI agents via self-custodial wallets and crypto rails, bypassing traditional KYC identity requirements to let agents pay for API resources and digital goods.

Brian Armstrong predicts that the agentic economy will eventually surpass the human economy in size, as autonomous AI agents outnumber humans in the near future.

Brian Armstrong describes Coinbase's internal AI harness, Toshi, which achieves recursive self-improvement by writing human code review feedback back into repository-specific AI brains.

Brian Armstrong reports that 88% of Coinbase's revenue comes from non-Bitcoin trading, highlighting a structural shift toward a broader digital asset economy.

Brian Armstrong states that Coinbase launched a tokenized stock product outside the United States to target the 4 billion global individuals who lack access to traditional brokerage accounts.

Brian Armstrong states that Coinbase's prediction market product reached a $100 million revenue run rate within months of launching, growing at over 100% quarter-over-quarter.

Brian Armstrong co-founded New Limit to pursue epigenetic reprogramming, using machine learning to identify transcription factors that can restore youthful cell function without changing cell identity.

Brian Armstrong announces that New Limit has successfully reprogrammed a human cell type in animal models, with Phase 1 clinical trials for alcoholic liver disease launching next year.

Brian Armstrong estimates that New Limit's initial liver disease drug alone could address a $20 billion market, though the ultimate goal is systemic rejuvenation of healthy individuals.

Brian Armstrong notes that Pew Research data shows 80% of Americans support embryo editing for disease prevention, predicting that genetic screening will eventually become a standard parental safety practice.

Brian Armstrong advocates for the creation of special economic zones or freedom cities on US federal land to bypass restrictive regulations for testing experimental energy, drone, and medical technologies.

Arm shifted from licensing individual IP components to delivering complete compute subsystems, culminating in its first physical chip, the ARM AGI CPU. This evolution addresses customer demand for faster market delivery and bypasses traditional chip design bottlenecks.

Rene Haas reports that 80% to 90% of Arm engineers use AI tools daily to accelerate verification, validation, and debugging. These backend processes represent the longest phases of the typical 24 to 36 month chip design cycle.

Rene Haas predicts that AI will enable automated chip design from conceptual idea to a final fabrication GDS2 file within five to ten years. This level of automation will initially apply to simpler, standard chip architectures.

Rene Haas projects that hardware supply chains will remain highly constrained for at least three to five years. The intensive compute and memory demands of transformer based AI models will continue to bottleneck access to wafers, substrates, and advanced packaging.

Rene Haas advocates for building more semiconductor fabrication facilities on United States soil to protect national security. He argues that domestic manufacturing clusters drive critical regional ecosystems, bringing in secondary industries like liquid cooling and energy infrastructure.

Rene Haas forecasts that general purpose robotics will replace human labor across construction, service, and security sectors. Arm microprocessors will power these machines, providing real time sensing at the limbs and managing central perception systems.

Rene Haas argues that CPUs remain essential in the AI era to orchestrate and route the tokens generated by accelerators during inference. Accelerators generate raw computational output, but microprocessors act as the systems decision makers to direct those outputs to users.

Google acquired bankrupt Spirit Airlines' data asset for $10 million to train AI models. Ofir Ehrlich notes that Mercor was a competing bidder, signaling a rising trend of companies buying defunct corporate databases to secure real-world training sets.

Ofir Ehrlich argues that synthetic datasets are insufficient for building functional agents. AI labs actively target real-world enterprise databases, such as public annual corporate reports, to understand true business hierarchy, operations, and communication flows.

Ofir Ehrlich explains that enterprise data remains locked inside disparate business units and unmonitored servers. Business leaders resist sharing this history with central data teams due to concerns over exposing sensitive information like executive compensation.

Gonen Stein states that security threats are shifting from human actors like ransomware to non-human AI agents with legitimate network credentials. These agents can drop tables or alter databases at extreme velocities, making immediate recovery systems essential.

Ofir Ehrlich warns that non-technical employees are using low-code tools to build custom AI agents. These shadow agents operate outside official organizational security guardrails, constantly transferring sensitive company data without IT oversight.

Ofir Ehrlich claims that the phase of token maximization is over as companies demand direct financial returns. Organizations are migrating to structured data foundations to limit token usage and lower the high cost of querying historical archives.

Ofir Ehrlich notes that Databricks is actively building its own databases and agentic workflows. The company is pivoting because it recognizes that an increasing share of modern enterprise data is now being generated by autonomous agents.

Ofir Ehrlich observes that enterprise AI adoption is shifting toward a forward-deployed engineering model. Tech providers are embedding their own engineers into legacy organizations to bypass long sales cycles and accelerate initial software deployment.

Ofir Ehrlich highlights a trend where investment firms purchase legacy enterprises with the explicit strategy of modernizing them into AI-first companies. This approach allows slow-moving businesses to improve profit margins and bypass traditional software procurement.

Max Hodak asserts the brain is literally a physical computer. Computational problems are solved by arranging matter in specific ways and allowing natural physical laws to execute state transitions.

Science received European marketing approval for the Prima retinal prosthesis, an implant that restores form vision by projecting laser images. Science acquired the underlying technology from French developer Pixium in late 2022.