什么火了
ARCHIVE近 30 天历史柱高表示当天的文章数
10/16—11/14 有历史数据
- Jensen Huang's resignation as CEO of Nvidia leaves CFO Colette Kress atop the Central Bank of AI : $42b in 2025-2026 M&A & tech buyouts, $61b of cash invested in equity across 229 private startups & 13 public companies, & $164.5b in lease guarantees & cloud capacity backstops. Last quarter Nvidia deployed 112% of operating cash flow into equity & M&A.
- Public disclosures from CoreWeave, Anthropic & wholesale data center landlords piece together the AI gross profit value chain in dollars per watt per year. At a 40% application margin, applications (40%), neoclouds (28%) & inference vendors (24%) capture 92% of total gross profit before GPU depreciation.
- KC, Chief People Officer at OpenAI & former CIO at Palantir, argues that efficiency is the wrong frame for AI. The prize is a reimagined firm where every role gains an agent manager & EQ becomes what separates people.
- US AI demand grows from 66gw in 2027 while tokens per watt compound 2.9x a year, so the grid only binds if token demand outruns efficiency. Break-even real token growth is ~2.5x a year with no cancellations.
- The cost of an autonomous 32-step corporate break-in fell 99% in five months from $7,500 on Claude Mythos Preview in April 2026 to $90 on open-weight GLM-5.3. Attackers run an expected-value calculation on AI compute ($1m of compute for a 1% shot at a $1b payout yields $10m in expected value), while defenders must maximize security effectiveness across speed (responding in minutes or seconds), scale (parallel agent swarms), & breadth (82-to-1 non-human identities).
- US AI capacity grows from 38gw today to an 81gw ceiling in 2027. Because 40% of compute serves inference & new Blackwell & Rubin racks lift blended monetization from $9.4b to $15.1b per inference gigawatt, theoretical US inference revenue rises 3.4x from $0.14t to $0.49t.
- Anthropic's Form S-1/A rolls financials through Q3 2026 ($17.0b revenue, +$8.6b operating cash flow) & reveals that $5.0b of short-term burst compute at $7 to $13 per GPU-hour lifted blended GPU-hour costs 82% in two quarters, flipping Q2's $974m operating profit into a $567m Q3 loss & driving $895.4b of long-term obligations.
- AI consists of four layers: application, harness & context, model, & data center. Each layer faces an unresolved economic question, & the four layers interplay directly: rising GPU financing costs increase the value of the harness, while model & inference co-design multiplies how many users the same data center hardware can serve.
- Two CEOs have set $1t revenue targets on their businesses while cloud growth accelerates at $30b to $50b quarterly scale. The first $10t company will be the one that crosses the trillion dollar chasm before AI erodes its legacy profit pool.
- Arvind KC, Chief People Officer at OpenAI & former CIO at Palantir, joins Office Hours to discuss the future of the people function. Why efficiency is the wrong frame when model capability doubles every six months. How the people team democratizes AI by amplifying a company's most valuable capability rather than rebuilding the whole org. How span of control shifts from managing seven people to directing hundreds of agents. & why emotional intelligence becomes the leader's edge when intelligence
- Coding & customer support were the first two AI markets. Attacks on banks, certificate authorities & core open source libraries have never been greater, suggesting security is third. A survey of 75 enterprise security practitioners finds security spend divides across five categories : security operations, SIEM, penetration testing, threat intelligence & vulnerability management. Every one carries significant AI token consumption. Greg Brockman argues security spend should vastly outstrip softwar
- Analyze Anthropic's S-1 IPO metrics : $4.6B revenue growing 11.9-fold, a derived 40% gross margin, $974M of operating profit in Q2, & $637B of compute commitments.
- Every major build-out has two phases per Carlota Perez: infrastructure, then the applications that pay for it. The telecom internet build-out ran capex twelve times ahead of application revenue at the 2000 peak, & took seven years to get application revenue above 10% of capex, & ten to reach 48%. AI has reached that 10% mark in four years, roughly half the time.
- Bardia Shahali, VP of Sales at Granola, joins Office Hours to explain how AI-era sales teams run smaller because revenue per rep multiplies, why comp plans moved from annual to quarterly, & why the sales pod splits by layer between infrastructure & application companies.
- SpaceX's third-quarter report & DeepSeek's inference disclosure together price one GPU card-hour four ways, from DeepSeek's theoretical $38 maximum to SpaceX's realized $2.44. Combined with a payback claim of under a year & a 2027 supply pipeline that is largely unbuilt, the data suggests large-scale inference is a healthy business with room for prices to climb.
- Anthropic is considered an AI company, but it should be valued as a cloud. On that basis, at $1.5t to $2t it would be worth roughly what Wall Street's estimates imply for a standalone GCP, on two-thirds the revenue.
- Five frontier models now sit within five points of each other in terms of intelligence, but their cost per unit differs by 3.8x. Google, which has built chips alongside its models for more than 15 years, prices lowest; Microsoft & OpenAI are now chasing the same advantage, & pricing should keep deflating even as frontier margins hold.
- Theory's second annual AI in Practice Survey finds four results that surprised us : 43% rate security a top pain against the 10% who say its tooling delivers value, 45% of teams route queries across models, Cloudflare hosts 62% of agent sandboxes, & 48% now read production logs with coding agents.
- Bardia Shahali, VP of Sales at Granola, joins Office Hours to discuss what the sales organization should look like in the AI era, drawing on his experience scaling go-to-market at Fireworks AI and Granola.
- AI models have reached commodity status as prices fall 41% while token usage surges 50%. To win relative share, labs should partner & resell models, control the user interface where distribution becomes the moat, & capture routing data to reduce the cost of training future models.
- Jeanne DeWitt Grosser, COO of Vercel, walks through the construction of the company's inbound sales agent: one GTM engineer at 20% time, a prompt written by the top SDR, six weeks of human-in-the-loop QA, & the eventual migration from a 1,000-line prompt to 14 deterministic rules with the model reserved for judgment.
- AI inference is set to surpass the database market to become the most important market in software, reaching ~$350b by 2027. This shifts software unit economics: usage dwarfs platform fees, gross margins compress below classic SaaS norms, and infrastructure becomes the dominant component of COGS.
- Explores the contrast between organic 'growing' of software via vibe-coding and deterministic 'designing' via state machines and formal verification.
- Tokenized real-world assets are the fastest-growing market in tech, compounding at 53% a month versus 25% for Anthropic. Stocks & commodities trade on blockchains 24/7, IPO pricing happens weeks before listing, & Allium now powers Bloomberg's RWA terminal dashboard.
- Anthropic & OpenAI both re-rated their run rates in 2026 by segmenting : a mandatory enterprise repricing against an 80% price cut on the cheapest tier.
- GPU rental prices doubled in six months while inference prices kept falling. The reconciliation is efficiency : the hinge that turns scarce silicon into cheap intelligence.
- AI has removed the typing from software engineering & left the system design. DHH told Rails World that writing code by hand is no longer economically productive ; Artemis engineers merged 16 pull requests per day by August, & Lauren Tan ships 2,000 a month. The craft of fitting code together by hand is over ; what remains is designing the loop that decides whether the output is right. Donella Meadows's three properties of great system design, resilience, self-organization & hierarchy, describe
- AI prices are collapsing, but the competition is not at the frontier. Anthropic held the Opus line at $5 & $25 across five releases while OpenAI cut Luna 80% then 50% again, & open models now run a majority of token volume at an 86% discount to closed models. The most capable model, Fable 5.1, commanded only 3.7% of gateway spending in its first twelve days. Demand is a normal distribution with a fat middle, & as intelligence per dollar explodes the tokens may shift to commodity. Whether that ha
- Machine-native models replace human-facing text generation with zero-token typed execution, cutting inference costs by orders of magnitude for basic programming primitives.
- A UC Berkeley study finds the harness sets the price of an answer : GPT-5.6 Sol costs 71% less on Pi than on Claude Code, the same model returning the same result, & none of the 42 harness comparisons shows a statistically significant quality difference. Two startups bidding on the same $250k contract earn 38% & 75% gross margins depending on what they built around the model, & the cheaper one pays back its sales cost in half the time. Effective harnesses take a deep customer understanding, rele
- Jeanne DeWitt Grosser, COO of Vercel, reports 90% automation of inbound sales development, a support agent handling 93% of cases, a 32x ROI on the sales development agent, & a total annual infrastructure bill in the single-digit thousands. Vercel took its inbound SDR team from 10 to 1.25. The bottleneck is not the model, it is the codified workflow.
- Amodei asked the industry to pace itself, but nobody defined the word. Five camps answered : interpretability wants time to understand models, labor wants time for workers, the economic camp wants growth to service the debt, the geopolitical camp wants position against China, & the regulatory capture camp wants no new rules. Each priced the consequences of a pause & none named a speed. The one mechanism that could produce a number, a training compute threshold, was tried in 2023 at 10^26 FLOPS,
- OpenAI disclosed that its researchers now supervise 3.14 agent-workdays for every 8-hour human workday. Rather than making programmers three times smarter, software development is turning into a 24-hour factory where inference behaves like heavy tooling capex running multiple shifts.
- AI token consumption expands in three waves : chat, a single agent, & a meta-harness orchestrating many agents. Each wave multiplies consumption not through faster generation but through parallelization, & the second wave has already overtaken the first.
- Hyperscalers & data center operators will issue an estimated $4t in debt over the next five years to finance AI infrastructure. This credit expansion equals 286% of US commercial paper, 143% of global private credit, & 91% of the US municipal bond market, transforming AI infrastructure into a macroeconomic credit cycle.
- Meta's Muse Spark 1.3 prices inference at two levels : $1.25/$4.25 per million tokens for private data, & $0.10/$0.20 for training consent : a 92% price spread. Michael Mauboussin taught that market prices contain information about underlying expectations. This spread establishes a liquid clearing price of $1.24/m tokens for user & agent prompt data, allowing Meta to vertically integrate the AI data supply chain & acquire post-training data at pennies on the dollar compared to specialized data l
- Automating draft scaffolding does not shrink human editing time; it raises the output ceiling. A closed-loop taste flywheel cuts draft churn from 47 to 3 while line-level edits hold steady at a median of 136, shifting effort from structural triage to rhetorical precision. Across five years of writing, the quality floor surged: the 10th percentile gained twice as much as the 90th.
- Over three years of testing AI writing systems, one-shot prompt templates failed. The working architecture is a closed-loop taste flywheel where draft versions collapse from 47 to 3, yet line-level edits remain flat at 130 per post. AI does not just save time; it raises the ceiling of what the same effort produces. The real question is whether we accept the pre-AI baseline or push for the higher-quality output AI now makes possible, the way chess engines raised the ceiling of human play without
- The frontier AI market is segmenting from both ends of its supply chain. Rationing programs & export limits decide who may run the strongest models. Downstream, enterprises standardize on one or two named vendors, & products ship with a default model inside.
- Three months into running r2, an agentic AI harness, four lessons held : an inbox beats a task list, routing between local & cloud models is essential, self-healing works but surfaces more errors before it surfaces fewer, and even a no-huddle offense needs a quarterback calling the plays a human still has to make.
- AI model companies buy wholesale electricity by the megawatt and resell it as cognitive work. The unit economics of turning power into intelligence.
- NVIDIA's Q2 FY27 revenue reached $96b with a $108b Q3 guide, but hyperscale revenue grew only 13% sequentially against 25% for everyone else. To fund the buyers filling that gap, NVIDIA extended payment terms — DSO rose from 45 to 60 days & receivables hit $63b — & built a $581b stack of supply commitments, power guarantees, leases & $101b of equity in the AI startups & neoclouds buying its chips.
- NVIDIA's Q2 FY27 revenue reached $96b with a $108b Q3 guide, but hyperscale revenue grew only 13% sequentially against 25% for everyone else. To fund the buyers filling that gap, NVIDIA extended payment terms — DSO rose from 45 to 60 days & receivables hit $63b — & built a $581b stack of supply commitments, power guarantees, leases & $101b of equity in the AI startups & neoclouds buying its chips.
- When designing an agent, should its session run for your entire five-year company tenure or reset every day? Perpetual sessions suffer from context rot and security exposure. The winning architecture is a daily coordinator that resets every 24 hours, delegating to ephemeral specialists and persisting state to local files.
- When designing an agent, should its session run for your entire five-year company tenure or reset every day? Perpetual sessions suffer from context rot and security exposure. The winning architecture is a daily coordinator that resets every 24 hours, delegating to ephemeral specialists and persisting state to local files.
- AI hardware bottlenecks cascade in sequential multi-year waves from GPUs to memory, SSDs, CPUs, HDDs, and physical data center shells, driving facility buildout costs to $20b per gigawatt through the classic Bullwhip Effect.
- AI hardware bottlenecks cascade in sequential multi-year waves from GPUs to memory, SSDs, CPUs, HDDs, and physical data center shells, driving facility buildout costs to $20b per gigawatt through the classic Bullwhip Effect.
- Local models now answer 89% of everyday chat & reasoning queries as well as frontier models, & their efficiency per watt has improved 5.3x in two years.
- Local models now answer 89% of everyday chat & reasoning queries as well as frontier models, & their efficiency per watt has improved 5.3x in two years.
- Every software product has its own grammar: Figma has frames and components, Salesforce has accounts, contacts, and opportunities, CAD has tools and constraints. Capability and required expertise rise together, so the most powerful tools are the ones fewest people can operate. AI lets users express intent in English while an agent handles the application's grammar, illustrated by a non-technical founder who directed Codex to drive CAD software and produce a dress once previously unmanufacturable