All-In from Davos: Crypto, AI Compute, and Industrial Robotics

概览

This episode collects several Davos interviews centered on how crypto, AI infrastructure, and robotics are moving from speculative narratives into regulation, deployment, and ROI-focused business decisions.

Brian Armstrong argues that U.S. crypto policy has shifted toward clearer rules, with stablecoins, tokenization, prediction markets, and on-chain capital formation becoming major themes. Andrew Feldman explains why AI inference speed, power, memory, and geopolitical chip policy are now central constraints. Jake Loosararian closes by framing robotics as the data-gathering layer needed to make AI useful in heavy industry, defense, energy, and manufacturing.

Across the episode, Davos is described as moving away from ESG/DEI language and toward business execution, dealmaking, energy, national competitiveness, and measurable AI returns.

分段落总结

[00:37] Coinbase at Davos

[事实] Jason introduces Brian Armstrong from Coinbase at the World Economic Forum and frames the conversation around global crypto regulation. [事实] Armstrong says Coinbase is pursuing crypto market-structure legislation while also holding commercial meetings with banks and country leaders. [事实] He says five of the top 20 global systemically important banks are using Coinbase to build crypto infrastructure into their products.

[02:16] U.S. Crypto Policy Shift

[事实] Armstrong says the Biden administration, in his view, tried to unlawfully kill the crypto industry in America. [事实] He credits Donald Trump with campaigning on making the U.S. the crypto capital of the world and pushing for clearer rules. [事实] Armstrong says about 52 million Americans have used crypto and frames crypto regulation as both a consumer and global competitiveness issue. [推测] The discussion positions crypto policy as a strategic industry question rather than only a financial-compliance question.

[04:28] Stablecoins and Banks

[事实] Armstrong says most bank CEOs he met at Davos are leaning into crypto, with one top-10 global bank CEO calling crypto their number one priority. [事实] He says the GENIUS Act requires U.S.-regulated stablecoins to hold 100% of assets in short-term U.S. Treasuries. [事实] Armstrong contrasts this with fractional-reserve banking and says Coinbase’s stablecoin payments to users are structured as rewards, not interest. [事实] He says some bank trade groups are trying to undo parts of the GENIUS Act, which Coinbase considers a red line.

[10:27] USDC, Tether, and Listing Standards

[事实] Armstrong says Coinbase has a strong relationship with Circle, and USDC is the largest regulated stablecoin because it complies with U.S. and European frameworks. [事实] He says Coinbase does not have an exclusive arrangement with USDC and also supports other stablecoins, including Tether and PayPal’s stablecoin in certain ways. [事实] Armstrong says Tether has helped people in emerging markets with high-inflation local currencies, while also noting it is not currently compliant under the GENIUS Act in the U.S. [事实] Coinbase’s listing approach is based on minimum standards for cybersecurity, rug-pull risk, legality, disclosures, and compliance rather than investment recommendations.

[14:40] Top Crypto Trends

[事实] Armstrong identifies three major crypto trends: all assets coming on chain, prediction markets growing quickly, and stablecoin payments growing quickly. [事实] He says Coinbase is working with Kalshi, is talking to Polymarket, and could also list its own prediction markets. [事实] He describes the “everything exchange” as a future where assets beyond crypto, including equities, can trade on chain.

[15:54] Stablecoin Payments and Coinbase Products

[事实] Armstrong says the biggest stablecoin growth area last year was B2B cross-border payments. [事实] He says companies using cross-border payments often face seven-day waits and high foreign-exchange fees. [事实] Coinbase Business serves small and medium-sized companies with payments, invoicing, tax, and accounting tools. [事实] Coinbase Developer Platform is described as an AWS-like platform for wallets, trading, payments, staking, and financing.

[17:16] Tokenization and Private Markets

[事实] Armstrong says tokenizing private-company shares should happen with the company’s permission because of employee liquidity and vesting concerns. [事实] He says Coinbase is discussing easier on-chain capital formation with the SEC and supports broader paths to accredited-investor status. [事实] He predicts private fundraising and eventually public offerings could move on chain. [事实] Coinbase Tokenize is described as a product for tokenizing funds, real estate projects, and other financial products.

[20:52] Funds and Unbrokered Investors

[事实] Armstrong says tokenization can democratize access, increase demand, reduce back-office fees, and remove settlement risk through instant on-chain settlement. [事实] He says BlackRock and Apollo have publicly said they want to tokenize their products. [事实] Armstrong cites Coinbase research saying four billion adults are “unbrokered,” meaning they lack access to investment products. [事实] He says Coinbase has integrated an AI agent into its app to teach concepts like dollar-cost averaging and tax-loss harvesting.

[23:53] California, Talent Exit, and NewLimit

[事实] Jason and Armstrong discuss California taxes, housing costs, homelessness spending, and business flight. [事实] Armstrong says he read that billionaire departures have already created a negative $10 billion tax hole for California. [事实] Armstrong describes a “voice or exit” dilemma: trying to fix California from within or leaving and helping builders relocate elsewhere. [事实] Armstrong also discusses NewLimit, a biotech company focused on epigenetic reprogramming, and says its first drug candidate is likely to enter clinical trials next year.

[29:34] Davos, AI, and Coinbase’s Internal Agents

[事实] Armstrong and Jason say Davos has shifted from ESG and DEI themes toward business, dealmaking, and growth. [事实] Armstrong names crypto and AI as the two most important technology trends and says they will converge because AI agents need payments. [事实] He says Coinbase built an internal AI system connected to company data sources including Slack, Google Docs, Salesforce, and Confluence. [事实] Armstrong says he uses AI agents to surface what he may not know as CEO, including internal disagreements and how he spends his time.

[38:54] Cerebras and Wafer-Scale AI Chips

[事实] Jason introduces Andrew Feldman, CEO of Cerebras Systems, which builds wafer-scale AI chips for inference and training. [事实] Feldman says the Cerebras wafer-scale engine is 56 times larger than a B200 and contains four trillion transistors. [事实] He says Cerebras systems cost roughly $1 million to $1.5 million on premise, while cloud access can be rented by token, month, or year. [事实] Feldman says early Cerebras customers included national labs, the military, and pharma organizations training non-LLM AI models.

[43:16] Inference Speed and Deep Research

[事实] Feldman says deep research consumes enormous compute because one user task triggers cascades of many smaller queries. [事实] He argues that faster inference can change AI usage in kind, not only degree, comparing it to Netflix changing after broadband. [事实] Feldman says Cerebras powers Cognition’s coding engine and aims for near-zero latency so users stay in flow. [事实] He says OpenAI placed a major purchase order with Cerebras.

[49:02] Power as the AI Data-Center Constraint

[事实] Feldman says the OpenAI deal was announced as 750 megawatts. [事实] He explains that data centers are now discussed in terms of power rather than square footage because power is the limiting constraint. [事实] He says Cerebras is building cloud infrastructure for OpenAI and that the power will be delivered over several years.

[50:13] Energy, Cooling, and Local Communities

[事实] Feldman says the cheapest power is hydro, followed by natural gas, and mentions West Texas, Wyoming, the Caribbean, Guyana, and the Nordics as relevant energy locations. [事实] He says Cerebras systems are water-cooled and often use closed-loop systems where water transfers heat without being damaged. [事实] Feldman says some hyperscalers mishandled rural community relationships by increasing local power rates or demanding tax discounts. [事实] He argues data-center builders should act as good citizens through local investment, jobs, schools, and utility-cost protections.

[57:05] Nuclear, Space Data Centers, and AI Demand

[事实] Feldman says nuclear is clearly worth pursuing but is probably not the main data-center power source for the next three or four years. [事实] He says data centers in space are an interesting idea, but cooling in a vacuum, satellite communication, and latency remain hard problems. [事实] Feldman says he believes AI compute demand is still very early, with enterprise adoption still tiny and consumer usage likely to rise sharply. [推测] His view implies that current AI infrastructure buildout is more likely constrained by supply than by near-term demand saturation.

[61:25] Compute, Memory, and Transport

[事实] Feldman says computer architecture depends on three things: calculation speed, memory, and transport. [事实] He says GPUs have large memory capacity but slow memory access, creating a bottleneck for fast inference. [事实] He says the memory shortage is partly caused by demand signals jumping from months of orders to much longer commitments across the supply chain. [事实] Feldman expects the memory market to take about 18 months to digest the shortage, with prices staying high.

[65:02] AI Geopolitics and Chip Policy

[事实] Feldman says the U.S. is ahead in chipmaking because of dense technical talent around Santa Clara, while China is running hard to catch up. [事实] He says China has advanced in open models and has benefited from top-down decisions to modernize its grid and add power. [事实] Feldman says the previous U.S. administration made mistakes by restricting chips from allies such as the UAE and Saudi Arabia. [事实] He says Trump’s administration has improved policy around allies, energy, grid modernization, and AI research, though he is unsure about selling H100s to China.

[75:04] Employment and AI Displacement

[事实] Jason raises concerns about young-worker unemployment, large tech headcount flattening, and AI-first startups doing more with fewer people. [事实] Feldman says current reductions in middle management are not yet mainly caused by AI, but by SaaS tools and organizational flattening. [事实] He says AI displacement is coming later, when whole job categories become much more efficient. [事实] Jason and Feldman briefly discuss Scott Adams and Dilbert as a cultural reference point for corporate middle management.

[78:52] Gecko Robotics and Davos’s ROI Turn

[事实] Jason introduces Jake Loosararian, CEO and co-founder of Gecko Robotics, which builds purpose-built robots to inspect ships, bridges, and other infrastructure. [事实] Jake says Davos has shifted toward business, negotiation, and ROI from AI. [事实] He says the key AI gap for infrastructure-heavy industries is the data needed to turn AI into productivity gains.

[81:31] Defense and Industrial Bottlenecks

[事实] Jake says about 30% of Gecko’s business is defense. [事实] He says the company helps address manufacturing-speed problems and weld-quality bottlenecks in areas such as submarines and destroyer turnaround. [事实] He says Gecko’s technology has been associated with manufacturing-speed improvements discussed by Admiral Houston.

[82:49] Energy, Built-World Data, and Robots

[事实] Jake says energy and power companies have become Gecko’s biggest growth area. [事实] He says Gecko’s robots and sensors inspect metal and diagnose the health of built-world assets such as bridges, dams, refineries, and submarines. [事实] He says combining inspection data with operational data can help extend asset life, increase production, lower costs, and improve decisions. [推测] Gecko’s strategic claim is that industrial AI depends on proprietary real-world data, not only internet-scale text or video data.

[84:53] Robot Economics and Repair Roadmap

[事实] Jake says Gecko is not focused on humanoid robots and argues that household tasks such as folding laundry are low-ROI use cases. [事实] He says the higher-value robotics opportunity is in oil, gas, power, defense, and manufacturing. [事实] He describes Gecko as an application and operational layer that can connect field robots, sensor data, and AI models. [事实] Jake says the roadmap moves from identifying and monitoring problems toward taking repair or manufacturing actions, including automated welding.

[86:44] Humans in the Loop

[事实] Jake says humans will remain in the loop for some time, especially through supervision and teleoperation. [事实] He says robots can reduce hazardous human work hours and move people out of risky environments such as deep-sea welding or bridge climbing. [事实] He argues robotics can lower the barrier to industrial jobs that traditionally require thousands of hours of expertise. [事实] He says the robotics community needs to get robots into the field, fail fast, prototype quickly, and improve manufacturing.

[89:14] Pittsburgh, Industrial AI, and Overlooked Problems

[事实] Jake says Pittsburgh’s steel history shaped his view of industrial infrastructure as the foundation for economic gains. [事实] He compares robotics-collected datasets to the infrastructure needed for AI models to deliver returns in real-world sectors. [事实] He says Silicon Valley often lives inside an internet-focused bubble and has overlooked energy, metal, manufacturing, mining, and defense. [事实] He encourages startups to look where attention and conversation are not concentrated.

[92:09] Robot Foundation Models and Missing Data

[事实] Jason suggests that LLMs, vision models, and world models could let robots learn tasks like bricklaying from web data. [事实] Jake says this kind of capability may be closer than people think, roughly in a three-year timeframe. [事实] He says the industrial world lacks the video and data corpus available for consumer or internet tasks. [事实] Gecko collects valuable industrial data by solving paid inspection problems in places like large refineries, where humans currently gather data manually in risky conditions.

播客点评/总结

[推测] The episode’s main value is that it connects three usually separate conversations: crypto regulation, AI compute infrastructure, and robotics for heavy industry. The strongest through-line is that frontier technology is becoming less about demos and more about law, power, data, deployment, and ROI.

[推测] The crypto section is most useful for listeners tracking stablecoins, tokenization, prediction markets, and Coinbase’s regulatory posture. The AI infrastructure section is strongest on practical constraints: inference latency, power, memory, grid limits, cooling, and geopolitical chip strategy.

[推测] The limitation is that the interviews are strongly pro-business and pro-technology, with limited pushback on regulatory, labor, or environmental tradeoffs beyond what guests themselves choose to address. The episode is best suited for listeners interested in tech strategy, venture investing, AI infrastructure, crypto market structure, and industrial automation.