Summary

Coinbase says the U.S. crypto regime has flipped from attempted extinction to institutional deployment. Brian Armstrong argues the Biden administration tried to “unlawfully kill this industry,” while Trump kept his promise to pursue a U.S. “crypto capital of the world.” Five of the top 20 global banks now use Coinbase infrastructure—including disclosed integrations with JPMorgan and PNC—while BlackRock wants to tokenize every fund.
The stablecoin contest is now a fight over Treasury economics, deposits and whether banks can reopen legislation passed four months earlier. Under the GENIUS Act, regulated stablecoins must hold 100% reserves in short-term Treasuries—roughly a 30-day maximum maturity, Armstrong believed—while Coinbase can distribute rewards when customers also trade, make payments or subscribe to Coinbase One. Armstrong says Coinbase passes customers “about 100% of the economics” and that bank trade groups he believes are trying to undo the law represent a “red line.”
Armstrong’s three leading crypto trends are the “everything exchange,” prediction markets and stablecoin payments. Equities and other assets are moving on-chain; Coinbase currently works with Kalshi, is talking with Polymarket and could operate its own prediction markets. The clearest stablecoin product-market fit is already B2B cross-border settlement, replacing seven-day transfers and high FX fees; demand for Coinbase Business is strong enough to create an onboarding backlog.
Tokenization’s biggest payoff may be cheaper private-market formation and liquidity, provided issuers retain control. Armstrong says private-company tokenization should require the company’s permission, because vesting and illiquidity can retain employees. He expects both fundraising and eventual public listings to move fully on-chain. Coinbase Tokenize targets funds and real estate, while Armstrong frames four billion “unbrokered” adults as the latent market for $100 or $1,000 allocations currently denied access to high-quality assets.
Crypto and AI converge when autonomous agents need native wallets and programmable money. Armstrong expects agents to use stablecoins because traditional finance assumes a human behind each product; inside Coinbase, an AI connected to Slack, Google Docs, Salesforce and other systems already surfaces hidden disagreements and audits his time allocation. His preferred mode is “reverse prompting”: asking the system what he should notice or how he could become a better CEO.
Cerebras is betting that inference latency, not merely model quality, will determine AI usage and market share. The displayed wafer-scale engine was described on-air as containing 4 trillion transistors and being 56 times larger than a B200, and Feldman says Cerebras aims to collapse multi-stage deep research from minutes to seconds—a “fundamental change in kind,” like broadband turning Netflix from a DVD service into a studio. OpenAI’s announced 750-megawatt Cerebras cloud order makes power delivery, rather than chip count or floor space, the operative unit of capacity.
Feldman sees no near-term AI-compute glut, but he does see an 18-month memory digestion and an unresolved geopolitical race. Consumer usage could rise from six or eight queries daily to 100, while every request consumes more inference; simultaneously, inflated 18-month orders have scrambled memory-demand signals and kept memory prices high, with HBM demand adding pressure. China remains behind in high-speed chips but ahead in open models and grid buildout, creating a recursive race where “by getting ahead, you get further ahead.”
Gecko Robotics argues that the highest-return AI opportunity is the physical economy, where usable training data barely exists. Jake Loosararian says defense is roughly 30% of Gecko’s business, with Admiral Houston cited as reporting manufacturing-speed improvements as high as 90%, while energy customers use robots to extend asset life and increase output. The roadmap runs from inspection to automated repair and welding, with skilled humans supervising fleets; generic bricklaying may arrive in roughly three years, but industrial autonomy requires proprietary data gathered inside refineries, shipyards and power plants.

Deep dive

1. Coinbase has moved from regulatory defense to bank infrastructure

Armstrong’s political assessment is deliberately unqualified: the Biden administration tried to “unlawfully kill this industry in America,” whereas Trump campaigned on making America the “crypto capital of the world” and then pursued clear rules. The constituency is no longer niche—Armstrong cited 52 million Americans who have used crypto.
His competitiveness case extends beyond domestic politics: roughly 500 million people have used crypto globally, Bitcoin was “the best performing asset class of the last decade,” China plans to pay interest on its central-bank digital currency, and major stablecoin issuers remain offshore. Repatriating that capital is therefore both financial policy and industrial strategy.
Commercial adoption is already concrete. Five of the top 20 global banks use Coinbase to build crypto products, with JPMorgan and PNC publicly named; another top-10 bank CEO called crypto the institution’s “number one priority” and “existential.” Coinbase also powers BlackRock integrations as the asset manager pursues tokenization across its funds.

2. GENIUS turns stablecoins into a direct challenge to deposit economics

The GENIUS Act requires U.S.-regulated stablecoins to hold 100% of their backing in short-term U.S. Treasuries, which Armstrong understood to mean maturities no longer than about 30 days. His safety shorthand: users are effectively betting that “the United States government is not going to fail in 30 days.”
Armstrong contrasted that structure with fractional-reserve banking, where deposits are lent out and runs remain possible. Calacanis supplied the visceral example: during the Silicon Valley Bank run, he forced a portfolio company to withdraw half its cash so it could make payroll, despite directors’ reluctance to abandon a 30-year banking relationship.
Coinbase’s payout is legally a rewards program, not interest. Rewards cannot depend solely on the balance: customers must also make payments, trade or subscribe to Coinbase One. When they qualify, Armstrong said Coinbase passes through “about 100% of the economics,” making the Treasury yield a customer-acquisition and retention engine.
Calacanis framed crypto companies as disruptive technology competitors to banks; Armstrong softened that to “mostly collaborative,” saying some banks are nervous while others are leaning in. He was firmer about trade groups he believes are trying to revisit GENIUS four months after passage: preserving the law is “a red line,” even if banks and crypto companies can still both win.

3. Coinbase wants an everything exchange, not a ratings agency

USDC is Coinbase’s largest regulated stablecoin relationship: Armstrong described it as compliant with GENIUS in the U.S. and MiCA in Europe. But Circle is not exclusive—Coinbase also supports PayPal’s stablecoin, supports Tether differently across jurisdictions and remains open to additional stablecoins.
Armstrong’s Tether view preserved both sides. He credited its distribution for giving people facing “70, 100% inflation year-over-year” access to dollars and said its team had “done a lot of good for the world.” Yet, in his understanding, Tether does not currently satisfy GENIUS requirements for 100% reserves in short-term Treasuries, leaving users to assess that distinction themselves.
Coinbase applies minimum listing standards around cybersecurity, developer rug risk, legality and compliance, then lets customers choose. Armstrong was skeptical of investment ratings, saying the organizations behind them felt politicized, and compared Coinbase with an “everything store.” A customer-rating experiment failed because token holders simply “talked their book,” leaving disclosures and baseline screening as the current model.
Prediction markets fit that open architecture. Coinbase works with Kalshi, is considering other providers, is talking with Polymarket and could list its own markets; Armstrong emphasized that none of those routes must be exclusive. The platform thesis is broader than crypto tokens: “all assets are coming on-chain for trading.”

4. Cross-border business payments are stablecoins’ first breakout workflow

Armstrong ranked the three fastest-moving crypto themes plainly: the everything exchange, prediction markets “growing like crazy” and stablecoin payments “growing like crazy.” The first expands the asset universe; the latter two create frequent transactional activity rather than relying solely on speculative token turnover.
The strongest stablecoin growth during the prior year came from B2B cross-border payments. Armstrong’s example was a merchant buying goods in Asia or Europe to sell in Brazil: the traditional route can impose seven-day delays plus substantial FX charges, while stablecoins compress settlement time and friction.
Coinbase Business packages those rails for small and midsize companies through payments, invoicing, tax and accounting tools. Armstrong said customers are “beating a path to our door,” producing a substantial onboarding backlog and a need to staff the team faster—useful evidence that demand is operational, not merely conceptual.
Coinbase Developer Platform is the infrastructure counterpart, described as “kind of like AWS” for wallets, trading, payments, staking and financing. Asked when consumers will settle poker games in stablecoins, Armstrong did not offer a date; he redirected to the measurable adoption already occurring in cross-border commerce.

5. On-chain capital formation attacks private-market friction and gatekeeping

Armstrong’s guardrail is issuer consent. A private company may deliberately use vesting and illiquidity to retain employees and align a team through an eventual exit; letting employees sell after one year could undermine that mechanism. Tokenization should therefore happen “with the permission of the companies.”
He expects the SEC conversation to progress from registered on-chain securities for accredited investors toward broader eligibility and eventually fully on-chain IPOs. Restricting private investments to wealthy people acts like “a regressive tax”—the investors who can already qualify capture appreciation before public buyers receive access.
Calacanis described today’s workaround as a “boiler room”: special-purpose vehicles raise from dentists and other high-net-worth individuals before locating the desired SpaceX or similar shares, sometimes charging a 10% load-in fee. Long private-company lifecycles then leave public investors absorbing a multiyear valuation “indigestion period.”
Coinbase Tokenize targets funds, real estate and other products, reducing back-office costs and settlement risk through instant on-chain transfers. Armstrong cited BlackRock and Apollo as firms pursuing broad tokenization, then widened the addressable market to four billion “unbrokered” adults who might invest $100 or $1,000 but currently can earn only through labor.

6. Armstrong is reallocating both personal capital and geographic loyalty

California prompted Armstrong’s “voice or exit” dilemma. He still loves the state, yet likened repeated tax and policy proposals to “an abusive relationship.” Once a builder leaves, the incentive can flip from repairing California to persuading other talent and companies to resettle somewhere more welcoming.
Calacanis estimated roughly 20% of California billionaires may already have departed and cited a projected $10 billion tax hole; Armstrong did not independently validate those figures. His operational evidence was clearer: California recruits demand higher compensation because housing and schooling can double their living costs, while San Francisco’s revenue-based tax was punitive enough to help drive Stripe away.
After Coinbase went public in 2021, Armstrong kept the CEO role but used some liquidity to fund “big bets” in the “world of atoms, not bits.” That led to NewLimit, a longevity company pursuing epigenetic reprogramming intended to restore functions human cells possessed when younger.
He originally expected roughly five years of basic research. Instead, NewLimit demonstrated human-cell reprogramming within its first two or three years, and Armstrong said its first drug candidate would probably enter clinical trials the following year—an unusually compressed timeline, but still explicitly framed as probable rather than guaranteed.

7. Davos has pivoted toward growth as AI becomes Coinbase’s oracle

Armstrong felt Davos had shifted away from global-government, ESG and DEI themes toward business execution, partly because of Larry Fink’s leadership and Trump’s agenda. Calacanis cited 5.6% GDP growth, 4.6% unemployment and roughly 2.8%-2.9% inflation; his prescription was deregulation, low-cost energy, clear rules and private-market building.
His technology through-line joins the episode’s two dominant sectors: “crypto and AI are the two most important technology trends,” and they will converge because agents must pay for work. Traditional finance assumes a known human behind each product; Armstrong expects agents instead to operate crypto wallets and settle in stablecoins.
On employment, Armstrong remained a “techno optimist.” Agriculture’s workforce share fell from roughly 80% in the early 1900s to around 3%, replacing backbreaking labor with previously unimaginable work; AI and robots may similarly create a transition period but ultimately produce abundance and new occupations. Loosararian’s example suggested robotaxis could arrive in roughly six years rather than decades.
Coinbase already hosts an internal model connected to Slack, Google Docs, Salesforce and other systems. Through “reverse prompting,” it flags undisclosed strategic disagreements, compares Armstrong’s stated time priorities with actual allocation—32% versus a desired 20% in one example—and answers questions such as what he changed his mind about most.

8. Cerebras built a giant chip before the giant workload arrived

The displayed wafer-scale engine was described on-air as containing 4 trillion transistors, roughly 56 times larger than a B200. The host also quoted configured on-premise systems at about $1 million to $1.5 million; cloud access ranges from roughly $0.50 to several dollars per million tokens, with monthly and annual rental options also available.
The first unit cost Cerebras about $500 million to create. Feldman’s founding wager was not that AI’s exact scale was knowable, but that it imposed a new computational problem and justified solving a 75-year-old architectural challenge: build a chip large enough to deliver 20x or 50x gains, not incremental 2x improvements.
Before language models existed, Cerebras found customers in national laboratories, the military and pharmaceuticals. Workloads included protein-sequencing models, vision and tasks at the boundary between high-performance computing and AI—the early demand that let the company refine a machine designed specifically for the later inference explosion.

9. Speed changes AI products, not merely their benchmark scores

Deep research is among today’s largest compute consumers because one task launches many threads, each spawning further queries whose outputs become inputs downstream. With perhaps 20 queries each triggering another 20, traditional 10-to-20-second responses create a “giant waterfall of time and answers.”
Calacanis’s research dossier once required an assistant roughly 16 hours, fell to eight hours with early AI assistance and now takes the model under 10 minutes. Feldman’s target is seconds: Cerebras aims to return those cascading results in four or 10 seconds, eliminating the “grab a cup of coffee” interruption.
His Netflix analogy captured the product effect: faster internet did not improve DVD delivery; it turned Netflix into a studio. Likewise, near-zero latency lets Cognition users remain “in the flow” while coding. Even imperceptible milliseconds matter, because waiting pushes users toward another service—as Calacanis does by launching the same query simultaneously across Gemini, Claude and ChatGPT.
Cerebras’s announced OpenAI agreement covers 750 megawatts delivered through its cloud over several years. Feldman explained why capacity is now quoted in power rather than chips or square footage: electricity delivery is the binding constraint, so megawatts express what the deployment can actually support.

10. Power abundance requires better siting, community economics and patience

Feldman ranked hydro as the world’s cheapest power, followed by natural gas in locations such as West Texas and Wyoming. Flare gas is especially attractive because petroleum operations once burned it as waste before Bitcoin miners demonstrated an alternative use; geothermal offers another regional option in the Nordics.
Cerebras systems use water cooling in a closed loop: water passes the chip, absorbs heat, is chilled and returns. Feldman rejected the idea that this inherently consumes or damages vast water supplies—cooler input helps, but the liquid is recycled rather than chemically contaminated.
He conceded that hyperscalers created legitimate local backlash by negotiating poorly and letting utilities spread infrastructure costs across residents for 20 or 30 years. A better compact would guarantee no rate increases, create construction work and fund schools as “a rounding error.” Home batteries could further absorb cheap surplus power and discharge during peak demand.
Nuclear is “obviously the right thing to do,” though Feldman does not expect new reactors to supply most data centers within three or four years. Space-based compute is further out: solar is compelling, but vacuum cooling, satellite communication and returning data to Earth remain real engineering problems. His estimate was eight to 10 years; Armstrong suspected somewhat sooner.

11. AI demand is early, while memory supply is temporarily distorted

Feldman rejected the overbuild thesis because enterprise workflow adoption remains tiny and even heavy consumers use AI only six or eight times daily. Usage could reach 100 interactions per day, devices will act autonomously, every engineer may gain a coding copilot, models keep improving and each interaction itself consumes more compute.
A computer must balance three functions: calculating, storing results and delivering them through I/O. Accelerating only one creates a bottleneck—“it doesn’t matter how fast the car can go if it can’t turn.” Cerebras’s senior architects therefore allocate power and silicon across computation, memory capacity, memory speed and I/O simultaneously.
Feldman argued that GPUs hold substantial memory capacity but access it too slowly for fast inference. He connected that weakness to the reported $20 billion purchase of Groq: “fast inference needs fast access to memory,” and the incumbent architecture lacked a sufficient answer.
The broader memory shortage partly reflects confused ordering signals. Buyers moved from roughly six months of demand to a full year, received uncertain delivery dates, then submitted 18-month forecasts; manufacturing output had not changed, but apparent demand exploded. Feldman expects about 18 months of digestion and sustained high prices, amplified by GPUs consuming large volumes of HBM, a DRAM variant.

12. The AI race spans chips, grids, standards and organizational design

Feldman sees the U.S. well ahead in chipmaking because several of the world’s best teams cluster around Santa Clara and improve by repeatedly building high-speed chips. China is running hard but remains behind there; conversely, it has moved ahead in open models and grid expansion through top-down power investment.
That asymmetry matters because AI is recursive: better tools make developers, knowledge workers and biotech researchers more productive, accelerating the next iteration. “By getting ahead, you get further ahead,” so initially small performance differences can become winner-take-all advantages.
Loosararian praised the Trump administration for better engaging allies such as the UAE and Saudi Arabia, improving CFIUS engagement and seeking consistent data-center rules. Letting allies build on American technology strengthens the U.S. standard; Feldman remained unsure about permitting H100 sales to China, calling it a genuinely difficult, non-clear-cut question.
On employment, Feldman agreed large AI displacement is coming but rejected AI as the primary cause of current cuts. Better SaaS already lets leaders oversee wider spans, reducing middle management’s role in moving information; companies are also flattening bloated post-hiring-boom structures. Later, AI will make entire categories “vastly more efficient” and require fewer people.

13. Gecko makes robotics answer to barrels, kilowatts and ship readiness

Loosararian said Davos CEOs have moved beyond generic AI enthusiasm to a harder question: “Where’s the ROI from all the AI?” For infrastructure owners, the missing ingredient is high-quality physical data. Gecko’s thesis, established before the current AI wave, is to make companies robot-native first and then layer models onto the resulting information.
Defense represents about 30% of Gecko’s business. Its robots inspect welds and manufacturing quality for submarines and help shorten destroyer turnaround; Loosararian cited Admiral Houston discussing manufacturing-speed improvements of up to 90%. The underlying constraint is a century-old industrial base competing with China’s manufacturing velocity.
Energy is the fastest-growing segment because asset health can be translated into production economics. By combining robotic inspection with operational sensor data, customers can judge whether to extend equipment life, push an asset harder, make more barrels per day or generate more kilowatts at lower cost.
Loosararian’s design rule is that robots should solve the customer’s fundamental business problem—not showcase dexterity. Whether the output is a barrel, kilowatt or ship released from dry dock, hazardous hours reduced and failures caught are “easy to underwrite.” Folding laundry, by contrast, is a weak early use case for an expensive humanoid.

14. Proprietary physical data is Gecko’s moat and the path from inspection to repair

Gecko’s roadmap begins by mapping the health of bridges, dams, refineries, submarines and other structures, then determining the correct intervention. Inspection results can feed automated welding, verify each weld and eventually train a “foundation model for welding”—closing the loop from identifying a defect to repairing it.
Humans remain essential as supervisors and domain experts who understand the consequences of one action versus another. Loosararian expects one welder to oversee perhaps 10 robots, with teleoperation moving hazardous work from bridges, deep water and industrial sites into air-conditioned control rooms.
Automation can also widen access to scarce trades. Gecko’s robots reduce the 10,000-hour learning barrier for inspectors and welders; Loosararian imagines training a Home Depot employee within months to operate systems safely and earn $100,000-$150,000, while the experienced worker supplies judgment and training data.
Generic humanoids may learn bricklaying from online videos within roughly three years, Loosararian estimated. Industrial work is harder because the necessary corpus does not exist: Gecko’s robots gather fused-sensor data while inspecting facilities, including from 100 feet up in a refinery. Gecko acquires that missing physical dataset while solving paid inspection problems, creating proprietary information for future autonomy.

Episode 170 - January 23, 2026 Captaining IMO Gold, Deep Think, On-Policy RL, Feeling the AGI in Singapore — Yi Tay