A curated summary of contemporary frontier discussions.
WEDNESDAY EDITION
Wednesday, October 21, 2026 · 16 episodes · 15 min read
AI computing power becomes a traded commodity
The Trends
14 showsAI is squeezing traditional software profits and changing how engineers get paidWhat's the story
Christina Cacioppo, CEO of security software firm Vanta, said on Invest Like the Best that corporate buyers immediately reject extra per-use fees for AI features. Software companies must absorb those computing costs into flat subscription prices, which shrinks their profit margins unless the tools keep customers from canceling.
On Empire, venture investor Michael Dempsey of Compound warned that open-source AI programs now copy standard business software on demand, while AI model usage grows tenfold a year. He now puts 70% of his fund into physical science and biotech startups because software tools without proprietary data or workflows will lose their pricing power.
Alex Lieberman and CJ Hess of AI engineering firm Tenex told Latent Space that 90% of big companies give staff AI chatbots, but only 5% run automated AI workers outside engineering. Tenex pays its coders for completed work instead of hours, creating a 4x pay gap between its fastest and slowest developers.
What's next
Gross profit margins at public software companies: if margins dip while subscription prices stay flat, vendors are eating rising AI computing bills without gaining pricing power.
Enterprise software renewal rates: if companies start replacing single-purpose apps with general AI assistants like Town's email tool, mid-sized software vendors will see customer cancellations spike.
23 showsCrypto money shifts from startups to tradable tokens, and traditional finance pushes backWhat's the story
Arthur Cheong of DeFiance Capital said on Forward Guidance that funds trading liquid tokens (crypto assets anyone can buy or sell on an exchange) returned 20% to 100% this year while Bitcoin lagged. Because struggling tokens drop 50% every six months, traders profit by betting against losers and backing winners like Hyperliquid.
On Empire, venture investor Michael Dempsey said crypto startup funding has gone dormant, pushing money toward discounted shares in established firms like Chainalysis and tokens that collect real user fees. For new crypto startups, that means fewer investors willing to back an idea before it has users.
The rise of offshore trading venues is drawing fire from traditional finance: CME Group CEO Terry Duffy argued on Odd Lots that US regulators must bar venues like Hyperliquid unless they verify users' identities. Offering 250x borrowed money without that check, he said, sets off forced-selling spirals and breaks US trading rules.
What's next
CFTC enforcement actions against perpetual futures (crypto contracts with no expiration date): reclassifying them as swaps would block US traders from offshore venues and hand volume back to CME.
Secondary market prices for locked venture tokens: if three-month boom-and-bust cycles continue, venture funds holding multi-year lockups will have to mark down their portfolios.
32 showsTraders are turning AI computing hours into a financial commodity like oilWhat's the story
On Odd Lots, DRW trading firm founder Don Wilson said AI cloud provider CoreWeave must offer bond buyers a 15% return because it rents chips to unrated customers. That slows data center building, so Wilson is creating futures contracts, pending US approval, that let cloud providers lock in future revenue and borrow more cheaply.
Evan Conrad, CEO of chip-time marketplace SF Compute, told Latent Space that traders rent Nvidia B300 chips for $5 to $6 an hour and resell that power running AI models for $25 to $40. Customers on his exchange also resell unused hours to recover $25 of every $100 spent, which lowers software builders' computing bills.
DRW is designing a contract, Wilson said on Odd Lots, that tracks the gap between raw chip rental costs and what AI companies charge per word generated. Oil refiners use a similar contract to lock in the price gap between crude oil and gasoline, so AI firms could protect themselves from sudden price drops.
What's next
The Commodity Futures Trading Commission's post-election ruling on GPU futures: approval lets Wall Street banks finance AI data centers at lower interest rates, speeding up construction.
Resale prices for older Nvidia chips on SF Compute: if buyers only trade the newest hardware, cloud providers holding three-to-five-year leases on older servers could face heavy losses.
Guests · Emmanuel AmeisenAI interpretability researcher at Anthropic
Language models do not just blindly guess the next word. Anthropic interpretability researcher Emmanuel Ameisen reveals that peering inside models like Claude shows they form internal world models, plan lines ahead, and hold abstract concepts across formats. Ameisen argues treating models like human actors rather than traditional programs is actually the most accurate way to understand their behavior.
Five things worth knowing
1Internal intermediate calculations prove models plan multi-step outputs in advance; Claude selects target rhyme words in its hidden layers before generating the preceding lines.
2Models form unified internal concepts across mediums; Claude activates the exact same circuit whether it processes a photograph of an eye, a vector drawing, or descriptive prose.
3Sudden capability jumps occur when smooth scaling crosses practical utility thresholds, so a model feels revolutionary overnight simply by moving from 80% to 95% task accuracy.
4Anthropic researchers treat alignment like biology instead of programming, directly monitoring internal mathematical activations to detect hidden deception before a model generates text.
5Caveat: Inspecting internal numbers cannot catch every failure mode because researchers understand only a fraction of the model's intermediate states, so deceptive behaviors can still remain hidden.
“a lot of what we do feels so much closer to biology than to engineering”Lenny's Podcast · 22:02
ConnectsAs frontier AI development resembles biology more than traditional coding, safety increasingly hinges on empirical neuroscience inside networks rather than written software rules.
◆
No. 02 · Latent Space
The Programmable Data Center — Evan Conrad, SF Compute
Guests · Evan ConradCo-founder and CEO of SF Compute
Evan Conrad, CEO of trading exchange SF Compute, argues that Nvidia and major cloud providers cannot build a true computing market because they would lose lucrative direct revenue. Instead, financializing AI compute requires neutral market makers who handle messy physical delivery, letting buyers resell unused GPU hours and hedge price swings across three-year contracts.
Five things worth knowing
1Traders can buy Nvidia B300 chips for $5 to $6 an hour and resell that power running AI models for $25 to $40 an hour, creating massive arbitrage.
2Customers who buy processing time through SF Compute recover an average of $25 for every $100 spent by reselling spare capacity to others on the order book.
3Compute contracts now stretch out three to five years, matching data center financing schedules and creating forward price curves similar to power and commodities markets.
4Delivery requirements vary wildly: batch computing jobs tolerate interruptions on an hour's notice, while live user-facing applications require one to three months of advance notice to move workloads.
5Caveat: Financializing compute ahead of standardized physical delivery creates extreme basis risk, where synthetic indexes detach from the actual cost of securing working chips in data centers.
“You have to kill your business in order to make this.”Latent Space · 34:20
ConnectsAs artificial intelligence chips become standard capital assets, computing power is shifting from software-style subscriptions into a physical commodity market driven by hedging, leases, and power grid constraints.
◆
No. 03 · The Cognitive Revolution
A World Model of Life: Biohub's Alex Rives on Virtual Cells, Measurement, & the $1.8B Data Push
Guests · Alex RivesHead of Science at the Chan Zuckerberg Biohub Network
Current virtual cell models fail because measurement tools cannot observe molecular activity at scale, says Chan Zuckerberg Biohub head of science Alex Rives. Backing a $1.8B multi-institution initiative, Rives argues predictive cell models require purpose-built hardware like laser phase plates and multi-scale data collection across billions of cells rather than standard transformer text architectures.
Five things worth knowing
1The $1.8B Virtual Biology Initiative unites Biohub, federal agencies, Meta, and DeepMind because predictive cell foundation models require mapping hundreds of billions of cells across temporal and molecular dimensions.
2Standard transformers fail across multi-scale biological structures, so models must instead learn probabilistic transition functions between cell states over time.
3Biohub deployed a laser phase plate for electron tomography to resolve single macromolecules in situ and replace the noisy proxies that stall drug discovery.
4Biological reasoning models cannot scale on web text alone; empirical laboratory verification loops must supply the physical ground truth for reinforcement learning.
5Caveat: Prioritizing molecular cell mechanics over longitudinal patient health records risks surfacing cellular targets that fail to alter clinical outcomes in whole humans.
“The things that we most want to know about biology, they're not represented in text.”The Cognitive Revolution · 54:02
ConnectsAI progress in the physical sciences increasingly shifts value from model weights to proprietary instrumentation and automated wet labs capable of generating high-resolution empirical data.
Markets & Macro
No. 04 · ChinaTalk
The US Can Stop China's Import Loophole | Nick Brown
Guests · Nicholas BrownCo-founder of the Center for Technology and Statecraft
Nicholas Brown argues the proposed MATCH Act would cement US compute dominance by cutting off ASML’s immersion DUV tool shipments to China. Because China cannot indigenize lithography tooling at scale before 2035, denying tool exports and servicing caps domestic AI chip output regardless of progress in etch or deposition.
Five things worth knowing
1China imported roughly 90 immersion DUV systems across 2024 and 2025. Those ASML scanners account for nearly two-thirds of the Dutch supplier's immersion shipments.
2A single ASML Twinscan NXT:1980 scanner can yield over 100,000 Huawei Ascend AI chips annually. China's current tool inventory already enables substantial domestic advanced fabrication.
3Even under scenarios where SMIC indigenizes etch and deposition, lithography throughput remains the binding limit. China cannot independently build advanced litho scanners at volume this decade.
4Denying maintenance contracts on deployed scanners degrades fab yields within 6 to 12 months. Subsystems like light sources eventually fail without specialized Dutch or Japanese field engineers.
5Caveat: The 150-day diplomatic window in the MATCH Act invites aggressive Chinese stockpiling. ASML and Tokyo Electron could front-load tool deliveries before restrictions take effect.
“It is a button that lets you win the AI competition, at least as measured by chipmaking.”ChinaTalk · 52:01
ConnectsExport controls are shifting from finished accelerator chips to capital equipment choke points, forcing allied governments to sacrifice tool vendor revenues to preserve a compute gap.
◆
No. 05 · Odd Lots
CME's Terry Duffy on the Disconcerting State of Finance Regulation
CME Group CEO Terry Duffy insists the CFTC broke regulatory discipline by approving 2,500 consecutive self-certifications. Novel prediction markets breach statutory anti-manipulation standards, he argues. Defending CME's derivatives franchise against encroaching crypto venues, Duffy demands regulators reclassify perpetual futures as swaps and bar offshore platforms like Hyperliquid unless they enforce strict KYC and clearing mandates.
Five things worth knowing
1The CFTC approved all 2,500 self-certifications under the current administration, but novelty contracts on hot dog eating or speech teleprompters breach Core Principle 3 against manipulation.
2Perpetual contracts are legally swaps rather than futures, because periodic funding rate exchanges match Dodd-Frank swap definitions and evade Section 1256 tax treatment.
3Offshore venues like Hyperliquid cannot legally enter the US without KYC, because offering 250x margin without participant identification causes liquidation cascades that violate Core Principle 9.
4Continuous 24/7 trading cannot scale to equities until commercial banking operates non-stop, as clearinghouses cannot pre-fund a $76 trillion market when bank payment rails close on weekends.
5Caveat: Duffy attacks retail prediction markets as manipulative gambling, but CME simultaneously monetizes sports via team indexes with Futureproof while exploring 24/7 commodity trading.
“If you can have an exit strategy, you will be more successful as you enter into anything.”Odd Lots · 35:19
ConnectsIncumbent financial exchanges are weaponizing post-2008 clearing and compliance statutes to slow down retail prediction venues and decentralized perpetual protocols trying to onshore.
◆
No. 06 · Odd Lots
Don Wilson on the GPU Market's Trillion Dollar Question
Hedging compute risk will unlock cheaper debt for neoclouds facing double-digit borrowing costs. DRW founder Don Wilson is building GPU futures to let cloud providers lock in forward capacity. Wilson argues the market already mirrors commodities, paving the way for cleared perpetual contracts and token spreads that financialize AI infrastructure.
Five things worth knowing
1CoreWeave's 15% bond yields show neoclouds face heavy capital costs when selling to unrated clients. Listed GPU futures allow them to hedge on-demand sales and secure cheaper financing.
2DRW is modeling the 'crunch spread' between raw compute costs and model inference tokens. This structure mirrors oil crack spreads by quantifying the processing margin of open-source models.
3The CFTC extended its review of compute futures past the November election. Wilson expects clearing approval because Silicon Data's benchmark tracks transparent, on-demand neocloud pricing under IOSCO standards.
4Perpetual contracts mechanically pass hard-to-borrow stock premiums directly to retail traders. Traditional brokerage accounts routinely confiscate these yields, leaving short-financing windfalls solely in dealer hands.
5Caveat: GPU hardware quickly becomes obsolete across multi-year cycles. Liquid forward curves might concentrate only on the latest architecture, leaving older clusters illiquid.
“in at at DRW we call this the crunch spread.”Odd Lots · 19:35
ConnectsCommoditizing compute capacity transforms data centers into physical commodity processors, shifting AI infrastructure funding from speculative venture equity to traditional debt and derivatives markets.
Crypto & Digital Assets
No. 07 · Unchained
'Bunker Mode': Is Crypto's Cryptography Running Out of Time?
Guests · Alex PrudenCo-founder and CEO of Project Eleven · Conor DeeganCo-founder and CTO of Project Eleven
Alex Pruden and Conor Deegan of Project Eleven, pitching their new Strongpoint custody product, warn that AI-assisted math solvers threaten elliptic curve cryptography faster than quantum hardware. Blockchains must design crypto-agile architectures supporting multiple signature schemes simultaneously, rather than retreating into rigid hash-based bunker modes or betting on single NIST standards.
Five things worth knowing
1Frontier AI models are breaking structured post-quantum schemes like Hawk and McEliece today, proving that algorithmic math problems fall to machine intelligence well before quantum hardware matures.
2Hash-based digital signatures avoid vulnerable algebraic structures entirely, but their larger payload sizes and slower verification impose severe bandwidth and throughput penalties on global blockchains.
3Privacy-preserving protocols like Zcash face existential exposure if public-key encryption fails; while signature schemes can rely on hashes, interactive zero-knowledge proofs and shared secrets require rich mathematical trapdoors.
4Social rollbacks and pre-deployed validator kill switches cannot replace cryptographic security, as coordinate-level resets destroy permissionless finality across decentralized networks.
5Caveat: Hash functions have historically suffered far more practical breaks and deprecations than elliptic curves, making hasty retreats to hash-only primitives an unproven risk hedge.
“we need to assume anything could break.”Unchained · 53:27
ConnectsAI acceleration shrinks the migration window for internet and financial security, forcing blockchains to design runtime agility before mathematical hardness assumptions collapse.
◆
No. 08 · Unchained
Could AI Break Cryptography Soon? Yes, but 'Bunker Mode' May Not Be the Solution
Guests · Alex PrudenCo-founder and CEO of Project Eleven · Conor DeeganCo-founder and CTO of Project Eleven
AI models now crack experimental cryptographic schemes in months, potentially threatening blockchain math before quantum computers ever arrive. Project Eleven founders Alex Pruden and Conor Deegan, who build post-quantum custody, argue that blockchains must prepare for rolling algorithm failures. Total inaction poses far greater danger than choosing imperfect replacement defenses today.
Five things worth knowing
1AI models recently broke experimental post-quantum schemes like Hawk because machine learning discovers mathematical flaws decades faster than physical quantum hardware matures.
2Elliptic curve cryptography protects nearly all major blockchains, but its underlying mathematical structure leaves it vulnerable to both quantum algorithms and AI pattern detection.
3Hashing addresses shields exposed keys from automated pattern discovery, yet users remain vulnerable whenever they sign an outgoing transaction and reveal their public key to the public ledger.
4NIST prioritized lattice-based signatures over hash-based alternatives because hash schemes create bloated transaction sizes and cannot support complex multiparty custody protocols.
5Caveat: No production cryptographic algorithm powering Bitcoin or Ethereum has actually broken yet, so hasty protocol overhauls risk introducing fresh software bugs without hard evidence of immediate danger.
“We need to assume anything could break.”Unchained · 41:02
ConnectsAs frontier AI models accelerate automated math proofs, decentralized networks face a new operational test: adapt cryptographic foundations or risk systemic financial failure.
Guests · Arthur CheongFounder, CEO and CIO of DeFiance Capital · Rajiv Patel-O'ConnorGeneral Partner at Framework Ventures · SmacPartner at Compound VC · Felix JauvinHead of Content at Blockworks
Crypto hedge funds are crushing venture returns this cycle because liquid dispersion lets active pickers extract alpha while venture paper remains trapped. Arthur Cheong of DeFiance Capital argued that fundamental crypto assets are trending cleanly, enabling profitable long/short execution. For LPs, beating Bitcoin requires liquid flexibility rather than illiquid, slow-moving venture books.
Five things worth knowing
1Most liquid crypto hedge funds returned between 20% and 100% year-to-date while major assets like Bitcoin lagged, proving dispersion now rewards discretionary stock-pickers over passive beta.
2Hyperliquid maintained consensus positioning and multi-billion-dollar valuation expansion without suffering the mean-reversion that typically crushes widely agreed-upon trades in previous crypto cycles.
3Early-stage AI tools allow seed-stage crypto startups to cut headcounts drastically. Founders now deploy five-to-ten-million-dollar rounds toward compute and API credits rather than engineering payroll.
4Liquid allocators budget a strict 5% to 10% entry price variance when building core positions. Short-term execution noise matters less than getting size into high-conviction fundamental assets.
5Caveat: Venture allocators still face crowded exits because secondary market participants price thematic narrative shifts in three months rather than years, leaving illiquid books mispriced.
“Before, when a name become very consensus in crypto, it tend to, you know, not doing well and peak. But this is like, I think, and also obviously depends on the market cap, but Hype has been like a huge market cap coin for like quite a while, and and it's very consensus, and yet it's still continue to go up and outperform.”Forward Guidance · 21:08
ConnectsLiquid token picking displaces illiquid venture as the dominant alpha generator, accelerating market cycles while compressing startup capital requirements through AI-assisted development.
Guests · Arthur CheongFounder, CEO and CIO of DeFiance Capital · Rajiv Patel-O'ConnorGeneral Partner at Framework Ventures · SmacPartner at Compound VC
Crypto fund managers Arthur Cheong, Rajiv Patel-O'Connor, and Smac argue that active liquid investing now beats locked venture capital. Crypto cycles compress into three to six months, while failing projects bleed 50% every half-year. With retail abandoning mid-caps for speculative memecoins, managers profit by holding liquid winners and shorting decaying protocols.
Five things worth knowing
1Short positions in crypto now generate positive returns on their own. Struggling projects consistently drop 50% every six months, letting managers profit on both legs rather than just hedging beta.
2Retail traders have abandoned mid-cap tokens for sub-$50 million memecoins to chase 20x gains. That leaves institutional capital to bid up a concentrated circle of protocols with verified usage like Hyperliquid.
3Fund managers budget 5% to 10% price slippage when building liquid token positions. Treating liquid tokens like venture bets lets allocators ignore day-to-day volatility to chase large long-term outcomes.
4AI coding tools allow early-stage crypto startups to launch with smaller headcounts. A $5 million seed round now extends much further, shifting budget allocations from payroll toward computing power and APIs.
5Caveat: Automated AI agent payments do not necessarily need blockchains. Standard bank APIs can handle agent commerce unless an autonomous program goes rogue and requires uncensorable rail systems.
“when you think about it like, if a AI agent go rogue, is it going to use a bank API or fintech API, or is going to use crypto? I think it's definitely going to be crypto.”Forward Guidance · 08:58
ConnectsAs software development costs drop, crypto venture shifts from funding large engineering teams to financing hardware and buying publicly traded liquid tokens.
Venture & Capital
No. 11 · Empire
Which Businesses Will Survive AI? | Michael Dempsey
Guests · Michael DempseyManaging Partner at Compound
Compound Managing Partner Michael Dempsey argues foundation model labs will destroy most application-layer SaaS wrappers, making venture bets on software pricing power unviable. Dempsey, who sat out 16 months of deployment during the 2021 bubble to preserve capital, now directs 70% of his fund's capital into deeptech and bio where physical experimentation creates durable moats.
Five things worth knowing
1Model labs will absorb downstream application value because thin wrappers lack the proprietary workflows and physical data needed to prevent instant margin collapse.
2With inference growing tenfold annually, software businesses expanding sixfold trail market replacement rates and will suffer severe repricing once early subsidies dry up.
3Open-source agents now decompile enterprise SaaS on demand, meaning only deeply embedded collaborative canvases like Figma maintain defensible moats against automated software cloning.
4Crypto venture remains dormant, pushing capital toward discounted private forensic firms like Chainalysis and liquid tokens with real fees rather than fresh pre-product protocol rounds.
5Caveat: Wet-lab biology may not remain protected by tacit experimental knowledge if automated lab robotics and synthetic generation accelerate clinical validation faster than five-year venture funds can adapt.
“the only thing free in company building is thinking”Empire · 13:17
ConnectsLong-duration capital faces obsolescence unless managers reject short-term momentum to price the deflationary impact of foundation models on software margins and human labor.
◆
No. 12 · How I Invest (David Weisburd)
$2 Trillion Investor on AI Portfolio Risk, Finding Top Funds, and Preparing for a Downturn
Guests · Tim McCuskerChief Investment Officer at NEPC
Advising $2T across 400 institutional clients, NEPC CIO Tim McCusker argues institutional allocators obsess over peer benchmarking rather than actual liability matching. LPs trap themselves into consensus asset allocations. By tracking rivals to protect annual compensation bonuses, institutional investors systematically miss the contrarian liquidity provision that generates true outperformance during market crises.
Five things worth knowing
1CalSTRS captured its co-investment alpha through fee discounts across all GP offerings because deal-by-deal selection failed to beat manager averages.
2Mega pensions managing $100B cannot move returns with $50M commitments, which locks massive pools out of early venture and into large-scale buyouts.
3Over half of private funds NEPC underwrites are Funds I through IV, which forces allocators to accept early operational gaps to access top emerging talent.
4Current cash yields near 4% make portable alpha redundant, while memories of 2008 margin calls still deter allocators from derivative-backed equity overlays.
5Caveat: Institutional allocators surrender roughly half of gross profits to pod shops like Citadel, tolerating steep expenses only because multi-manager platforms effectively replaced bank trading desks.
“A long maybe is the worst outcome there for the GP.”How I Invest (David Weisburd) · 25:23
ConnectsInstitutional capital concentration in mega-cap tech and multi-manager pods reveals how career risk and liquidity constraints drive allocators away from contrarian rebalancing during market stress.
◆
No. 13 · Capital Allocators
Building a Berkshire for the Middle Market – Tracy Britt Cool (EP.521)
Guests · Tracy Britt CoolCo-founder of Kanbrick and former CEO of Pampered Chef
Turning around an underperforming business took four years instead of two because shifting culture and people is far harder than spreadsheet models suggest. Former Warren Buffett deputy Tracy Britt Cool, now pitching her investment firm Kanbrick, argued that private equity firms exit too quickly because traditional fund deadlines force sales before operational gains can compound.
Five things worth knowing
1Kanbrick uses a 15-plus year holding structure, which removes the pressure of the three-to-five-year private equity clock and lets management make decade-long investments.
2Cool built a free peer community of 4,000 executives, which directly sourced four of the firm's seven deals without competing in banker-run auction processes.
3At calibration firm JM Test, Cool quadrupled operating earnings over three years and returned all initial investor capital through cash flow without selling the company.
4Cool targets businesses earning $5 million to $50 million, abandoning sub-$5 million targets after finding smaller firms lack the management depth to support institutional operating systems.
5Caveat: Kanbrick promises founder-friendly stewardship, but turning struggling businesses around often requires replacing long-tenured family personnel and firing misaligned leaders.
“Everything is harder in operating than it seems.”Capital Allocators · 00:11
ConnectsAs higher interest rates squeeze debt-fueled buyout returns, private equity firms increasingly pitch operational turnarounds and longer holding periods to justify fees.
Builders
No. 14 · Invest Like the Best
Christina Cacioppo - Building an Enduring Software Company in the Age of AI - [Invest Like the Best, EP.496]
Guests · Christina CacioppoCEO and co-founder of compliance automation platform Vanta
Vanta founder and CEO Christina Cacioppo argues enterprise software buyers will instantly reject per-token markups on AI models. Instead of passing through inflated charges, software providers must absorb model costs directly into flat subscription fees. Survival depends on accumulating proprietary customer data that automated agents can query, not on owning proprietary underlying algorithms.
Five things worth knowing
1Enterprise buyers detect marked-up model fees immediately, so software vendors absorb computing costs directly to protect customer retention.
2Companies fine-tune AI models primarily for marketing to venture investors, since custom model weights rarely beat standard prompts on actual task quality.
3Top software compensation climbed tenfold over five years as tech companies paid escalating premiums for high-agency operators who independently solve messy problems.
4French software companies adopted Chinese AI models because US policy shifts proved that access to American commercial systems carries severe business continuity risk.
5Caveat: Absorbing AI compute costs compresses software profit margins today, which leaves vendors vulnerable if heavy agent usage fails to produce superior long-term customer retention.
“You get 2x revenue multiple if you say agent.”Invest Like the Best · 42:54
ConnectsAs foundational AI models become commoditized utilities, enterprise software value concentrates in proprietary internal company data and distribution networks rather than custom model training.
◆
No. 15 · a16z Podcast
Why a $1M Startup Wasn’t Worth Building in the Age of AI
Guests · Jean-Denis GrezeCEO and founder of Town, formerly CTO of Plaid · Alex RampellGeneral partner at Andreessen Horowitz
Tax prep seemed ripe for generative AI, but Town founder Jean-Denis Greze killed his tax startup after hitting $1 million in revenue. Incumbents move faster than annual sales cycles allow. Backed by investor Alex Rampell, Greze rebuilt Town around daily email and calendar workflows, betting that horizontal AI assistants will outcompete specialized tools.
Five things worth knowing
1Selling AI tax software failed because customers only buy once a year. Incumbents update their software faster than a startup can overcome the dead time between filing seasons.
2As token costs collapse, AI assistants can automate low-frequency, high-value tasks like comparing vendor price sheets quarterly. A dollar of computing can suddenly save a restaurant thousands in wholesale costs.
3Neutral software aggregators win by routing queries across competing models. Relying on a single provider leaves an enterprise stranded whenever a rival model pulls ahead in code or reasoning.
4AI agents break the classic ad-funded web covenant. Software bypasses promotional clutter to retrieve raw data directly, threatening revenue models for aggregators like Amazon that monetize human search attention.
5Caveat: Automated assistants require access to internal corporate communications to deliver real value, but strict data privacy rules and liability fears make enterprise security teams hesitant to permit full email scraping.
“The future looks like a lot less human in the loop than we see today.”a16z Podcast · 41:33
ConnectsAs AI agents consume web data without viewing ads, traditional ad-supported platforms will lose their core revenue stream and must find direct monetization models to survive.
◆
No. 16 · Latent Space
AI in NYC: How Tenex Builds an AI-Native Engineering Team — Alex Lieberman & CJ Hess
Guests · Alex LiebermanCo-founder of Tenex and executive chairman of Morning Brew · CJ HessSpecial projects lead at Tenex
Billing clients by the hour penalizes engineers who adopt artificial intelligence to work faster. Tenex co-founder Alex Lieberman and lead CJ Hess argue the fix is paying software developers strictly for output. Selling that model to big corporate clients exposes a gap: most enterprises have access to models, but lack the standardized processes needed to let automated workers run.
Five things worth knowing
1Tenex pays engineers by completed story points rather than hours, creating a 4x pay gap between its fastest and slowest software developers.
2Ninety percent of large enterprises gave staff access to frontier language models, but only five percent run automated agents for non-engineering work.
3CJ Hess found software developers cannot automate corporate workflows if the underlying process lacks documentation; companies must formalize standard operating procedures before agents can execute them.
4Plunging model inference prices ended enterprise fears over runaway token usage; corporate buyers now worry more about data permissioning and security controls across departments.
5Caveat: Measuring productivity through story points invites engineers to pad estimates or deliver shallow code that technically passes review without solving business problems.
“You can't automate a process that does not exist.”Latent Space · 23:25
ConnectsAs artificial intelligence collapses raw coding costs, consultancies and software firms must overhaul billable-hour business models or watch client fees evaporate.
Also aired, not covered
Tokenized Is the Stablecoin Sandwich Dead? · not processed in time
All-In Dylan Field: Is AI Replacing Designers?, Design is the New Code & Democratizing Creativity · not processed in time
Training Data (Sequoia) Inside the Robot Data Factory: Mecka's Josh Gao on Sensorizing Humans to Train Robots · not processed in time
a16z Podcast Should You Trust ChatGPT With Your Health? · not processed in time
20VC Who Wins in the Next Decade of VC | Margin and Price Matter Less in a World of AI | Sameer Dholakia · not processed in time
This Week in Startups Anthropic Gives Away 5% of Its Shares, Clone’s Muscle Robot, ElevenLabs Steals Podcaster's Voice · not processed in time
Not Enough Time · Wednesday, October 21, 2026Summaries are written with AI from the linked episodes and checked against them, but can contain mistakes. Follow the links to hear what was actually said. Nothing here is investment advice.