AI budgets tighten as agent bills outrun plans

Enterprise (5)

• Menlo’s enterprise AI estimates put annual spending at $33 billion in 2025 and nearly $177 billion in 2026; its reported growth multiples were 5x from 2023 to 2024 and 3.9x from 2024 to 2025. The survey covered about 400 U.S. enterprise decision makers responsible for roughly $4 billion in AI spend.
• Menlo found enterprise AI use rose from 55% to 65% in a year, while average spending per AI-using enterprise grew 4.7x. Whittemore reads the gap as evidence that agents let companies scale spending with work and parallel activity, not employee headcount.
• Companies spent about 2.5 times what they budgeted for AI: planned spending growth was 1.9x, but actual average spending rose 4.7x. Whittemore says seat-based budgets failed to anticipate token-heavy agent use, leaving CFOs to treat AI as a new spending category.
• AI-native vendors dominated coding, finance and operations, and design spending, while incumbents retained most other enterprise categories. IT was the clearest incumbent stronghold: established vendors captured 91% of IT AI-app spending, compared with 9% for AI-native companies.
• Menlo found SaaS application growth slowed from 13% in 2025 to 11% in 2026, but most new AI spending came from new budgets rather than existing software budgets. Whittemore warns that agents using software without conventional interfaces could still pressure SaaS.

Models (3)

• Enterprise buyers shifted spending toward more capable, more expensive frontier models: their share rose from 35% to 63%, while mid-tier, small, and older models lost share. Menlo also reported labs cutting small-model prices by more than 80%, showing that cost competition and demand for capability coexist.
• Anthropic gained enterprise API share from 12% in 2023 to 55% this year, as OpenAI recovered to 31% and Google fell to 9%. Whittemore links Anthropic’s lead to Claude Opus models becoming a default for long-running agentic coding tasks.
• OpenAI’s Images 2.6 added short animated GIF generation, frame editing, speed controls, and reference-image input. OpenAI said users create 4 billion images per week, up from 3 billion a couple of months earlier, and the model is 50% faster, generating images in as little as six seconds.

Open Source (1)

• Open models accounted for 4% of enterprise AI spending but 19% of tokens, up from 11% of tokens the prior year. Kimi captured 61% of open-source spending; Whittemore argues token share may better indicate where adoption is heading than revenue share.

Coding (1)

• Revenue for Anthropic’s Claude Code and OpenAI’s Codex rose from $1 billion at the end of 2025 to $24 billion at the end of 2026. Other coding platforms, including Cursor, Cognition, Factory, and Lovable, also grew from $2.6 billion to $7.8 billion.

Agents (3)

• Agentic coding applications grew 7.7x, just behind financial-services AI applications at 8x, though coding spending was far larger. Menlo estimated $34 billion in coding-app spending against $800 million for financial services; customer success, sales, legal, and marketing also expanded.
• Claude’s personal-finance feature can connect to bank accounts through read-only Plaid access and analyze spending for budgeting, but it will not provide investment advice. Anthropic says users should rely on licensed professionals and regulated firms for major financial decisions.
• Meta’s Muse agent added voice and video conversations with animated avatars, selectable voices, and subtle expressions, using Meta’s live-voice models and real-time avatar technology. The feature lacks visual input for reading nonverbal cues and is available on mobile, web, and Mac.

Markets (3)

• AI stocks suffered their worst drawdown since 2020, with the Magnificent Seven losing a combined $1 trillion in market value. The S&P 500 fell 4.4% and the Nasdaq 6.4%, while semiconductor suppliers Micron and SanDisk dropped more sharply than major platform companies.
• Whittemore says the broad selloff looked like forced deleveraging after semiconductor losses, rather than a sudden change in company fundamentals. Reports said Walleye returned capital, Lone Pine had collapsed, and Millennium withdrew $1.5 billion in leveraged capital from two outside managers.
• Anthropic’s rumored pre-Thanksgiving IPO became less likely as markets fell: prediction-market odds dropped to 60%, and odds of a listing by year-end fell to 66%. The company’s reported $1.6 trillion secondary-market valuation drew investor skepticism, with one IPO investor arguing for roughly $1 trillion.

Safety (3)

• Anthropic’s first independent safety report from Meter said the pace of model improvement roughly doubled over the past year, largely because researchers use AI assistance. Meter reviewed three unreleased models trained since July; the latest was completed in September or October.
• Meter warned that AI-driven research could produce a sudden emergence of extreme capabilities if systems begin accelerating further AI development autonomously. Ryan Greenblatt called the report a major update toward ASI within one to two years; Nathan Lambert questioned whether existing benchmarks can measure that pathway.
• Whittemore says Meter’s report leaves gaps as an independent audit: Anthropic redacted evidence, and the report did not assess sandboxing or other cybersecurity safeguards. Critics also questioned Meter’s closeness to Anthropic and argued that an evaluation focused on AI research acceleration misses other safety concerns.

Big Tech (2)

• OpenAI says ChatGPT now handles search requests at 20% of Google’s volume, although Adam Fry counts any query that triggers a web search. Sundar Pichai said AI Mode queries are slowing Google search revenue growth because they monetize less effectively.
• Adam Fry said ChatGPT ads have not hurt product usage, and that users’ limited backlash may encourage OpenAI to expand advertising. Nathaniel Whittemore argues that search growth could strengthen OpenAI’s advertising business.

Models (1)

• GPT 6.2 Astra and GPT 6.2 Sol appeared on Google Cloud’s backend, which Nathaniel Whittemore says usually signals an imminent release. The coming model comparison is whether GPT 6.2 can beat Opus 5.5.

Coding (2)

• Meta acquired Conductor, a startup that coordinates multiple coding agents running in parallel, to help build future coding products. Conductor founder Charlie Holtz teased a different approach to coding interfaces, while Meta’s MuseCode has drawn little attention.
• OpenAI agreed to acquire Pierre Computer Company, whose open-source tools help teams store codebases and track changes. Codex already uses Pierre’s tools, and the acquisition could support OpenAI’s ambition to add integrated services and compete with GitHub.

Agents (2)

• Meta’s and OpenAI’s coding acquisitions point to competition shifting toward agent products and the overall developer experience, rather than model quality alone. Nathaniel Whittemore says leading labs appear to be preparing for a market where frontier models become less differentiated.
• JPMorgan and Bank of America said they would block agents from accessing customer accounts, citing the risk of mistakes. Lead Bank and Increase Bank are testing agent-specific accounts, spending controls, emergency lockouts, and APIs through fintech partnerships.

Biology (1)

• Google’s drug-discovery spinoff Isomorphic Labs is reportedly raising $5 billion at a $50 billion valuation to expand its operations. The company builds on DeepMind’s AlphaFold work, and Demis Hassabis shifted to chairman partly to focus on scaling it.

Elections (1)

• Democrats’ election sweep put the Senate, House, and most state governments under Democratic control for two years, including Texas, where data-center construction has been central to the boom. Public resistance is strong: 43 of 44 data-center moratorium votes passed, and incoming Governor Gina Hinojosa is expected to be less supportive than Greg Abbott.

AI Infrastructure (5)

• Citrini Research argues that data-center construction faces compounding power, grid-connection, labor, supply, and permitting constraints, alongside turbine delays and rising input costs. New state policies may add municipal concessions, clean-energy requirements, ratepayer protections, and slower approvals, raising project costs without stopping construction outright.
• Nathaniel Whittemore argues that local leaders could allow data centers to proceed in exchange for transparency, full payment for infrastructure and power costs, environmental safeguards, and community benefits. Incoming Texas Senator James Tallarico has called for companies to pay for energy, recycle water, and obtain residents’ consent.
• Google, Meta, Amazon, and Microsoft all spent below their third-quarter capital-expenditure targets after construction delays. Microsoft attributed a $5 billion shortfall to a rented facility that was not ready, while Oracle warned it might invoke force majeure over a New Mexico project’s power-connection problems.
• Nathaniel Whittemore and market analysts argue that a slower build-out could help hyperscalers lower capital spending, protect margins, and preserve free cash flow while easing pressure on power, labor, and supply chains. Matt Farley says constrained data-center capacity could raise compute rents and make infrastructure scarcity a competitive moat.
• The Hill reported that hyperscalers are discussing a shared data-center framework with environmental groups, potentially giving Democrats enforceable community and environmental commitments while offering companies cover to moderate spending plans. Jessica Lesson says the market’s split reaction across hyperscalers, chips, and neoclouds suggests AI weakness need not pull the entire market down.

Markets (1)

• Investors reacted differently across the AI trade: the S&P 500 and Nasdaq rose while chip and neocloud stocks fell, and the Magnificent Seven gained $1 trillion in market value. Seema Shah warned that volatility would continue amid possible Iran-war escalation, a December shutdown, and Federal Reserve rate hikes.

Regulation (2)

• Progressives including Abdul Saeed and Yasemin Ansari have backed an AI pause, while Bernie Sanders has proposed jailing people who develop AGI. Nathaniel Whittemore considers a federal pause unlikely under a pro-AI president and a veto threat, and says Democrats remain divided between moderates and the Democratic Socialists of America.
• Nathaniel Whittemore expects more room for agreement on AI policy than the pause-versus-acceleration debate suggests, including possible liability for labs, safety-incident disclosures, and third-party audits. He argues that wider public debate is appropriate as model capabilities keep advancing.

Elections (2)

• Democrats won both chambers in the midterms, while voters in more than a dozen Ohio towns and in Kansas, Florida, and California backed restrictions on new data centers. Texas Governor-elect Gina Inohosa campaigned for a construction pause pending new regulations.
• The hosts say affordability, government corruption, and opposition to war drove the Democratic wins more than culture-war messaging. Jason expects many new Democratic officeholders to govern from the center, while the co-host says party leaders largely waited for Trump’s policies to damage his standing.

AI Infrastructure (5)

• Jason says Texas, Georgia, and Ohio together represent more than 40% of planned U.S. data center capacity. Bloomberg reported Microsoft was souring on neocloud providers, while Nscale lost a deal tied to a U.K. data center.
• Jason gives Donald Trump high marks for backing AI infrastructure, energy expansion, and growth, but says Trump made data centers a partisan wedge issue. Jason argues split government provides useful checks and balances after Republicans lost both chambers.
• Jason argues data centers can win local support by paying taxes, creating jobs, funding community benefits, and protecting water and electricity prices. He warns that slowing construction could trigger a recession lasting two to four years, and says operators must control noise, pollution, odors, and energy impacts.
• Jason says the Republican split over AI centers on job losses, wealth concentration, and fears that technology firms will leave workers in a permanent underclass. He connects those concerns to opposition from Steve Bannon, Nick Fuentes, Tucker Carlson, and Megyn Kelly.
• OpenAI pledged to cover the full cost of powering its data centers, solicit local input, publish construction plans, and avoid nondisclosure agreements with public officials. Jason says repeated, direct communication about utility bills, water use, taxes, and community impact is essential to building trust.

Markets (1)

• Jason says markets corrected 5% to 6% after the election, but sees the selloff as overblown if tariffs ease and the Iran war ends. He calls the decline a buying opportunity in companies including Nvidia, Amazon, Uber, Tesla, and SpaceX.

Regulation (2)

• Jason criticized Trump family meme coins soon after Trump took office, arguing that leaders responsible for fixing crypto regulation should not launch speculative tokens while doing so. He says government officials should keep personal trading at arm’s length from their public duties.
• Jason says taxing loans backed by appreciated assets could address wealthy people borrowing against holdings instead of selling them. He described taxing margin loans above a threshold, while noting the gap between ordinary income and long-term capital-gains rates.

Trade (1)

• Jason says Trump’s tariff goals were defensible but the rollout was chaotic, with public threats replacing a planned negotiating strategy. He argues leaders should hear directly from affected businesses; a clothing-company CEO described how tariffs were affecting his operations.

Models (2)

• Anthropic signed its first content licensing deal with the Associated Press for election coverage, with Claude expected to provide attributed, up-to-date polling information. Jason distinguishes licensing from unsettled questions about training on purchased content and using copyrighted material in AI outputs.
• Jason argues AI companies should share revenue with publishers and other rights holders rather than leave licensing disputes unresolved. He floated giving media businesses a portion of AI revenue, potentially through platform-wide arrangements that let publishers opt out.

Big Tech (2)

• Jason thinks Apple could let customers choose between cloud models and local models, using its own devices as an AI platform. The hosts also discussed an expensive touch-screen Mac configuration, with Jason questioning the value of touching a laptop display when an iPad can handle that use.
• The hosts debated whether technology executives would return to progressive messaging after Democratic victories. Jason says Democrats should invite business leaders into policy discussions, while warning that executives who shifted toward Trump may face distrust if they reverse course.

Agents (1)

• A host described using a Plaud pin and agent to route voice notes across Gmail, Calendar, Slack, Notion, Claude, and Grokbot, then build a product demo. Jason requested keyword-triggered actions and offline recording that syncs after a ski run, ideally through cellular connectivity.

Psychology (1)

• Jason says only a small share of casino gamblers tip, and he prefers putting dealers into a hand so they can share the outcome. He described a home poker game that sometimes includes the server in a bomb pot, giving the worker a chance at a large payout.

Society (1)

• The hosts traded Austin food opinions: one criticized Whataburger, while Jason said he would reconsider his view of P. Terry’s and praised Velvet Taco’s brisket and tikka masala tacos. The conversation framed these local chains as familiar parts of Austin life.

Media (2)

• Jason recommended Godzilla Minus One, praising its emotional connection to the creature and its postwar Japanese setting. The hosts discussed the follow-up, Godzilla Minus Zero, which moves the story into the 1950s and was playing in theaters that week.
• The co-host recommended the Korean monster film Hope, describing its long runtime, violent action, shifting story, and major twists. The same co-host praised season two of Apple TV+’s Dark Matter for treating travel between realities as a psychological problem rather than a superhero spectacle.

Agents (6)

• Jason Calacanis estimated his company spends about $40,000 a year on AI tools and gets more than $2 million in software value, including a syndicate backend, podcast infrastructure, and accelerator tools.
• Nick Harris said Lightmatter runs roughly 10 to 20 agents that track chip projects, manufacturing progress, blockers, and customer deadlines; he has spent $10,000 to $20,000 in a week building software and workflows.
• Jason Calacanis would pay $1,000 monthly for agents, while Peter Morales said he had spent $100,000 in a couple of weeks and would also pay at least that monthly; Nick Harris put his personal-use ceiling near $500.
• Peter Morales uses an hourly service to scan company Slack and email him customer-related issues, while avoiding employee DMs; he said operations staff have started building their own bots as adoption spreads.
• Jason Calacanis built a custom Austin home-search site by having agents find off-market listings, trace broker histories, and compile neighborhood reports; he said the work took under an hour across three or four weeks.
• Jason Calacanis said agents can turn repeated tasks into deployable software, build browser extensions, and remove sponsored Amazon listings; his unsponsored-results extension took six minutes and used an estimated 30,000 to 40,000 output tokens.

AI Infrastructure (7)

• Mark Jeffrey described Bittensor subnets as a distributed compute market: Liam supplies raw machines, while inference networks such as NGI, Gensyn, and Targon set up models and sell inference; he said NGI offered inference at one-tenth CoreWeave’s cost.
• Mark Jeffrey said Bittensor compute capacity can scale as demand raises miner incentives, though new supply takes time to come online; its marketplace mixes data centers with machines run by individual operators, including at home.
• Peter Morales argued that debt-funded neocloud expansion faces a squeeze, while Nick Harris expected AI prices to rise because lenders are risk-averse and providers cannot keep subsidizing heavy usage indefinitely.
• Peter Morales said his company is moving staff toward on-prem compute and self-hosted models, while Nick Harris expects higher prices and more efficient use of existing capacity as lenders retreat from risky infrastructure financing.
• Nick Harris said Lightmatter’s photonic interconnects can let a gigawatt data center train a model three times faster and produce twice as many inference tokens per second with the same power and GPU count.
• Jason Calacanis cited Elon Musk’s Colossus buildout at more than one million GPUs, with capacity growing from 300 megawatts to about one gigawatt and capital cost rising from $13 billion to $36 billion.
• Sam Altman argued that local models waste scarce DRAM because home memory is used only about five percent of the time; Nick Harris countered that AI’s memory needs will keep growing, invoking Jevons paradox and the limits of two-dimensional memory layouts.

Digital Sovereignty (1)

• Mark Jeffrey argued that locally run AI protects control over prompts and compute; he pointed to Bittensor subnet 95, which lets people run inference on home computers and rent excess capacity back to the network.

Safety (1)

• Peter Morales said formal methods teams still need human review for safety-critical code, while stronger reasoning models can tackle harder proofs; Code Metal applies those verification practices to defense, automotive, aerospace, and banking software.

Models (1)

• Nick Harris argued that mathematicians remain valuable because expertise helps them ask the right questions, even when AI can solve difficult problems; Peter Morales said AI-generated proofs still need scrutiny, and the Navier-Stokes episode raised concerns about credit and prompt use.

Elections (3)

• A Washington Post review found ChatGPT and Copilot relied heavily on three small election-information sites, including Ballot Brief; the review also found factual errors, while Mark Jeffrey said chatbot sourcing risks resemble the influence risks of search engines.
• Jason Calacanis and Peter Morales warned that low-quality or fabricated web pages can shape chatbot answers, while companies already market services that influence what ChatGPT says about businesses; Google previously blocked Gemini from answering election questions.
• The hosts questioned whether large donors or foreign governments could manipulate prediction markets with hidden bets, and noted that a self-described performance-art polling firm still entered poll aggregators, illustrating how easily bad inputs can spread.

Energy (1)

• Jason Calacanis predicted Texas data-center policy could shift after the election, arguing that Governor Greg Abbott’s temporary pause might end if the state can assure residents about water and power; another panelist said the industry depended on Texas expansion.