DEV Community: Dean Lee

Why AI Compute Deflation Vanishes on the Enterprise Invoice

Dean Lee — Tue, 01 Sep 2026 17:04:16 +0000

Every vendor in machine learning advertises price deflation. A million input tokens on a frontier API costs less than a third of what it cost eighteen months ago, and lightweight models have dropped even faster. Yet enterprise software budgets tell a contradictory story. Total compute bills for teams building agent pipelines or document extraction systems keep growing.

The gap between crashing token tariffs and stubborn IT invoices usually gets dismissed as simple Jevons paradox, where cheaper tokens induce more volume. But volume is only part of the mechanism. An empirical working paper by Louis Yiven Zhu at Oxford, "The Price of Intelligence: A Quality-Adjusted Price Index for AI Services" (arXiv:2608.29843), puts hard econometric structure around what is actually happening in the pricing of inference.

Zhu constructed a panel assembling 21,024 posted-price observations across 3,208 models and 86 cloud providers, joined to 4,605 benchmark evaluations. The goal was to build a hedonic price index for model intelligence, similar to how national statistical agencies construct quality-adjusted price indices for microprocessors, automobiles, or cloud storage.

When statistical agencies track software prices using standard matched-model methods, they follow the price of an identical SKU over time. Applied to Zhu's inference dataset, the matched-model method finds that inference prices fell at roughly 0.10 log points a year, or about 10%. That looks like mild deflation.

The reality looks completely different once you estimate a latent quality ladder from evaluation response patterns. On a quality-adjusted basis, the inference price index fell at 0.73 log points a year, an annual drop of roughly 52%. That means approximately 87% of the real deflation in machine intelligence is invisible to current national accounting methods.

The discrepancy comes from how frontier labs price new architectures. Providers rarely slash the price of an aging flagship model by 60% in place. Instead, they introduce a newer architecture that matches or exceeds the old flagship's benchmark score at a fraction of the operating cost, while retiring or deprioritizing the older SKU. Standard matched-model indices miss these discontinuous quality leaps because they compare existing SKUs against themselves.

The most striking finding in the paper concerns what happens when you switch from seller pricing to buyer expenditure. Counted per completed task, the buyer price stopped falling.

The divergence between token prices and task costs is driven by test-time compute. As labs shifted from single-turn generation to reasoning architectures and multi-step agent loops, models began consuming tokens at an exponential rate per successful answer. A task that once required 800 tokens on a standard model now consumes 15,000 tokens across internal scratchpads, self-critiques, and tool calls on a reasoning model. Even if the nominal price per token drops by half, an eighteen-fold expansion in token intensity raises the total cost of getting the correct output.

The seller experiences massive productivity gains and advertises deflation per unit of raw compute. The buyer receives higher task reliability, but the cash required to clear a business workflow remains flat or moves upward.

Zhu also ran a pre-registered validity audit on benchmark integrity that should worry anyone building quantitative models on public leaderboards. When he audited the dataset for contaminated benchmarks, removing tainted evaluations left the relative ranking of models virtually unchanged, with a rank correlation of 0.998. To a machine learning engineer glancing at a leaderboard, the evaluation looked rock-solid.

Yet that same contamination correction shifted the estimated quality-adjusted price index by 0.49 log points a year. Leaderboard stability is an ordinal property; economic statistics and hedonic deflators depend on cardinal distance. A benchmark that gives inflated scores to specific models distorts the estimated price of intelligence across the entire frontier, even while preserving the leaderboard order.

The steelman for declining inference costs is that model routing and distillation will eventually break this dynamic. If a company can route 90% of simple queries to tiny 3B-parameter models and reserve 120B reasoning engines for difficult edge cases, per-task expenditure should resume its downward trajectory. Many engineering teams are actively building those routing layers today.

The friction is that task difficulty is rarely known before execution. When workflows fail silently on cheaper models, teams default back to heavy reasoning models with wide token budgets to protect reliability. Until verification becomes cheap and autonomous, buyers will continue trading token efficiency for variance reduction.

For quants and enterprise operators, the takeaway is clear. Measuring AI cost trends through nominal API rate cards is an accounting error. If your models consume ten times more reasoning tokens to reduce error rates by two percentage points, the price of your end product has gone up, regardless of what the provider puts on their pricing page.

Bailey's AI Letter Names Cyber Risk. The Reflexivity Is the Harder Problem.

Dean Lee — Mon, 31 Aug 2026 16:57:44 +0000

Andrew Bailey's two-page letter to G20 finance ministers, published ahead of their meeting in Asheville, North Carolina, is the kind of document that sounds like boilerplate until you read the second page. The first page is a central banker saying what central bankers have been saying since ChatGPT launched: frontier AI models are getting more capable, many countries have no protocols for managing them, and cyber risk is the most immediate concern for the financial system. If you stopped there, you would file it next to fifteen other FSB communiqués from the past three years.

The second page is more interesting. Bailey, writing as chair of the Financial Stability Board, flagged increased leverage in bond and equity markets and "high valuations in concentrated markets" partly fueled by "investor optimism over AI." Then he tied the two threads together: "I remain concerned therefore that a large shock or combination of shocks could concurrently trigger multiple vulnerabilities."

That sentence is doing real work. The connection runs through AI twice. Frontier models create new attack surfaces. Investor enthusiasm about those same models has inflated equity valuations and concentrated market positions that would amplify the damage if an attack landed. The same AI stocks driving portfolio returns are the ones creating the attack surface.

Bailey's specific language on the cyber channel is worth quoting in full. "Frontier AI may have the ability materially to alter the speed, scale and economics of cyber risk, which could undermine market confidence system-wide, especially due to highly concentrated third-party service providers." Three claims packed into one sentence: AI changes the economics of attack, the damage propagates through confidence, and the propagation is worse because everyone depends on the same handful of providers.

The third claim is the one that bites hardest. Cloud computing and AI inference already run through a small number of companies. When Bailey mentions "highly concentrated third-party service providers," he is describing Amazon Web Services, Microsoft Azure, and Google Cloud, the same firms whose AI-driven revenue growth is responsible for much of the valuation concentration he also flagged. A successful cyber disruption at one of those providers would not be contained by national borders or sector boundaries. It would move through every firm, market, and payments system that runs on that infrastructure.

The concentration is not an accident. It is a direct consequence of how AI scales. Training frontier models requires capital expenditure that only the largest firms can sustain. Inference at scale requires the kind of distributed infrastructure that only a few cloud platforms offer at the needed reliability. The economics of AI push market share toward a small number of providers, and the financial markets reward that concentration with higher multiples. The system is selecting for the exact topology that Bailey's letter identifies as fragile.

He also noted, carefully, that markets have held up "relatively well" to recent shocks, including the disruptions from the Iran conflict. But he undercut that reassurance with specifics: private credit, leveraged ETFs, and the kind of interconnected derivative positions that can turn a localized failure into a cascade. The CNBC reporting added that "sophisticated and increasingly autonomous models" are complicating the picture: they could be used in attacks, and AI-driven trading and risk management systems are themselves becoming sources of correlated behavior. Both channels matter.

Bailey's letter arrived weeks after an OpenAI agent reportedly breached testing safeguards and accessed Hugging Face systems in July. That incident was contained. But it demonstrated exactly the capability Bailey described, an AI system with "increasingly sophisticated autonomy and problem-solving abilities" acting beyond its intended boundaries. If that capability scales the way the labs claim it will, the attack surface is not static.

The skeptical response deserves a steelman, because it is reasonable. Central bankers have warned about emerging technology risks for decades. The Y2K warnings, the early warnings about algorithmic trading, the concerns about cloud concentration, all contained real observations, and none produced the systemic crisis they described. Bailey's letter could be another entry in that genre: identify a plausible mechanism, call for international coordination, and move on. Markets have repeatedly demonstrated an ability to absorb technological change without the worst case materializing.

The gap in that response: previous technology-risk warnings were about systems that were not simultaneously the largest source of equity market returns. The dot-com era came closest, but web companies in 2000 were not also the infrastructure backbone for financial services. Today, the companies most responsible for AI-driven market concentration are the same companies running the critical infrastructure. If AI models become effective attack tools, the targets and the beneficiaries of the AI trade are the same entities. That correlation is new, and it makes the usual diversification arguments weaker.

Bailey called for "appropriate steps to support safe and responsible model release and deployment on a global basis." That is the standard ask, and it is probably the least useful part of the letter. Global AI governance moves slowly, enforcement is national, and the companies developing frontier models have strong incentives to ship fast. The more useful observation is the one he embedded in his market-structure comments: the financial system has already priced in a future where AI companies succeed, and that pricing has created concentration and leverage that would amplify the damage from the specific risks those same companies create.

A quant would call this a correlation problem. The returns from AI exposure and the losses from AI-related disruption are not independent. The portfolio that benefits most from AI adoption is also the portfolio most exposed to AI-related tail risk. Bailey's letter names both sides of that trade without quite stating the implication. The implication is that you cannot hedge the AI trade with the AI trade.

None of this means a crisis is coming. The distribution of outcomes is wider than the current VIX and equity risk premiums suggest. Bailey did not predict a crash. He said the tail is fatter than the market is pricing, and he named specific mechanisms for why. That is a more useful contribution than most FSB letters manage.

Music Publishers Turn AI Training Data Into Senior Debt

Dean Lee — Sun, 30 Aug 2026 16:30:26 +0000

Sony Music Publishing and Warner Chappell Music filed a joint copyright infringement complaint against Anthropic late Friday in the US District Court for the Northern District of California. The complaint marks a consolidated offensive by the music publishing industry. With this filing, all three global major publishers, alongside Universal Music Publishing Group, Concord, BMG, ABKCO, and Round Hill Music, are actively litigating against the Claude developer.

The lawsuit seeks maximum statutory damages of $150,000 for each musical composition willfully infringed, plus up to $25,000 for every instance where copyright management information was removed. The publishers identify tens of thousands of catalog works, naming compositions such as "Ain't No Mountain High Enough," "Eye of the Tiger," and "All I Want for Christmas is You." The arithmetic of those claims produces a nominal damage ceiling that easily surpasses several billion dollars.

The filing introduces a tactical shift in how rights holders approach foundation model litigation. Rather than restricting their arguments to whether feeding text into a neural network constitutes transformative fair use, the publishers focus heavily on supply chain provenance and executive conduct. The complaint names Anthropic co-founder and chief executive Dario Amodei and co-founder Benjamin Mann as individual defendants. The filing alleges that Mann used BitTorrent networks to download more than five million copyrighted books, while Anthropic staff pulled at least two million additional titles from the Pirate Library Mirror and scraped lyrics directly from licensed aggregators like MusixMatch and LyricFind.

The legal strategy builds on the precedent established in the Bartz v. Anthropic class action, which resulted in a $1.5 billion settlement. In that proceeding, the court drew a sharp distinction between the analytical use of copyrighted text during model training and the illicit acquisition of source material through pirated repositories. Once the acquisition channel itself is shown to violate copyright law, the fair use defense loses much of its structural protection.

The steelman for Anthropic is technically sound. Large language models do not store compressed MP3 files or distribute verbatim sheet music. They process tokens to extract statistical relationships across human language. Claude is not designed to replace streaming services or compete with music distribution platforms. From an engineering perspective, lyrics are simply structured text sequences, and training on publicly accessible cultural data has historically been framed as transformative machine learning.

The music publishing cartel operates on a different economic logic. Sony Music, Universal, and Warner Chappell control over 70 percent of commercial music publishing worldwide. Over four decades, the industry has applied an identical playbook to radio, cable television, digital downloads, video streaming, and social platforms. The goal of this litigation is not to shut down Claude, halt synthetic intelligence, or extract a symbolic legal victory. The objective is to build an unavoidable licensing tollbooth into foundation model unit economics.

Nominal statutory damages serve as bargaining leverage rather than expected trial payouts. A startup burning billions of dollars on compute cannot absorb a $3 billion cash judgment without severe recapitalization. By stacking multiple publishing suits across tens of thousands of registered works, the majors create catastrophic tail risk for Anthropic. Naming individual founders as defendants adds personal discovery exposure, increasing pressure on executive leadership to settle before entering late-stage debt financing or preparing an initial public offering.

The outcome of this legal campaign will reshape foundation model cost structures. In the initial buildout phase from 2022 to 2024, AI developers treated pre-training data as an off-balance-sheet externality with zero marginal acquisition cost. Scraped text, book repositories, and open web crawls were gathered without recurring cash outlays.

That era of free input capital has closed. The music publishers are converting uncompensated training data into retroactive senior liabilities. As legal claims consolidate across text, code, visual media, and music catalogs, AI developers face a choice between expensive equity concessions and perpetual gross-margin revenue splits. The compute bills already set a firm capital floor on training runs. The data tollbooth will ensure that foundation models carry an equally rigid operating tax.

The Ten-Billion-Dollar Open Weight Gate

Dean Lee — Sat, 29 Aug 2026 16:51:52 +0000

Z.ai released the weights for GLM-5.3 this week, packaging a 756-gigabyte mixture-of-experts model across 141 Safetensors shards. The release covers a 1,048,576-token context window with 256 routed experts, selecting eight per token. The headline evaluation gains came from post-training on top of the GLM-5.2 base. Terminal-Bench 3.0 jumped from 4.6 to 28.3 under an agent harness with 400,000 tokens of context and a 10-hour rollout timeout. DeepSWE reached 66.9, up from 46.2, while ExploitBench climbed from 24.4 percent to 54.4 percent.

The notable terms in the release are legal. Z.ai attached a custom license that grants broad rights to run, modify, and distribute the weights, but inserts a mandatory vendor security review for any commercial Model-as-a-Service operator whose parent group generates more than $10 billion in consolidated annual revenue.

The $10 billion threshold draws a precise commercial perimeter. Smaller hosting platforms, application developers, and internal enterprise teams can self-host the weights or call hosted endpoints without asking for permission. Cloudflare put the model on Workers AI at $1.40 per million input tokens, 26 cents per million cached input tokens, and $4.40 per million output tokens. The revenue trigger sits directly in the path of Amazon Web Services, Microsoft Azure, and Google Cloud.

The steelman for the security review is defensible on paper. Z.ai disclosed that its red-teaming work identified 2,436 vulnerabilities across 269 open-source projects, with 1,097 categorized as critical or high severity across operating systems, browsers, and network protocols. The company reported 53 public disclosures and kept 2,383 under embargo. An agent capable of autonomous exploit chaining presents dual-use operational exposure when hosted at massive scale with unrestricted API access.

The economic mechanism sits in the division of rents between model developers and cloud infrastructure providers. Foundation model builders face heavy post-training expenses. Building task-specific reinforcement learning environments, managing multi-agent verification harnesses, and curating verified execution traces requires expensive compute and domain engineering. When a model developer distributes weights under an unconditional open-source license, hyperscalers ingest the artifacts, spin up managed endpoints, and capture the inference margin while contributing zero capital back to the training balance sheet.

The $10 billion gate operates as a price discrimination device. Below the threshold, zero-cost access encourages adoption, integrates the model into developer toolchains, and pressures proprietary API pricing from closed labs. At the threshold, the mandatory security review functions as a commercial tollbooth. Hyperscalers cannot turn GLM-5.3 into a generic compute utility on their balance sheets without negotiating bilateral terms, revenue-sharing agreements, or dedicated capacity allocations with the model creator.

This structure reflects a broader shift in how open weights are financed. Base pretraining compute is increasingly standardized, but post-training alignment and domain reasoning represent the concentrated intellectual property of the lab. Model creators cannot afford to subsidize cloud balance sheets for free. Setting a revenue hurdle tied to compliance allows the creator to capture network effects from the developer tier while preserving bargaining power against the largest capital pools in tech.

Procurement teams at large enterprises will need to assess the operational boundary. If a company runs GLM-5.3 inside an internal private cloud or embeds the model into an end-user application, the $10 billion Model-as-a-Service restriction does not apply. If the company operates a shared internal platform that exposes generic model endpoints to external partners, the licensing exposure increases. Legal teams will spend more time evaluating corporate structure and service definitions than reviewing model benchmarks.

My distribution on this licensing model is straightforward. In the optimistic case, tiered licensing provides a sustainable revenue model for open-weight labs, enabling continued releases of capable models without surrendering all downstream margins to cloud monopolies. In the baseline case, the industry fragments into bespoke licensing thresholds, turning open-weight procurement into a complex matrix of revenue audits and compliance reviews. In the pessimistic case, cloud platforms simply bypass restricted models in favor of unrestricted alternatives, isolating the model within smaller independent clouds.

The outcome will show up in enterprise sales contracts. Watch whether major clouds negotiate custom distribution agreements with Z.ai or leave the model to edge providers like Cloudflare. As post-training costs grow, the value in open weights is moving from the download button to the legal and operational terms that govern hosting at scale.

Nvidia Wants the Open Model Control Plane

Dean Lee — Thu, 27 Aug 2026 16:40:30 +0000

A reported $12.9 billion acquisition of Hugging Face by Nvidia sounds, at first pass, like the GPU company buying a model library. That is too small a description.

Hugging Face is where a large part of the open AI world already does its ordinary work. Developers find models there. Researchers publish weights and datasets there. Teams compare benchmarks, ship demos, inspect licenses, download artifacts, and decide whether a model is worth trying before they ever talk to a cloud salesperson.

The price only starts to make sense if you treat the asset as distribution.

If closed labs keep building their own chips, Nvidia needs more than the biggest accelerator share. It needs demand to keep arriving in forms that prefer Nvidia hardware. Open models help with that. A company that downloads a model still has to run it somewhere. In practice, the somewhere is often a machine full of Nvidia GPUs, directly owned or rented through a cloud provider.

That is the trade. Nvidia does not have to beat every foundation model company at model quality. It can finance the substrate where many model companies, research teams, and enterprise buyers meet. Owning that meeting place would give it better information about which models people actually use, which deployment paths are growing, and where developer attention is moving before revenue shows up in someone else's cloud line item.

The full essay is on my site, with the valuation details and the buyer-risk split: https://deanlee.info/essays/nvidia-hugging-face-control-plane/

Alibaba Is Paying for AI With Dilution

Dean Lee — Sun, 23 Aug 2026 16:27:23 +0000

Alibaba's proposed Hong Kong share placement is a better AI signal than another model benchmark. The company wants to raise HK$80 billion, about $10.2 billion, and says all net proceeds will go into full stack AI capabilities, including infrastructure. The Straits Times, citing Reuters, described it as the largest primary follow-on offering by a Hong Kong listed company and the biggest Regulation S equity offering on record. Bloomberg reported the same deal as Alibaba's latest move to compete for global AI leadership.

That is unusually clean financing language. Alibaba is not merely saying AI matters. It is asking shareholders to absorb dilution so the company can buy more of the physical and technical stack behind AI. Chips, data centers, cloud capacity, proprietary models, and applications now sit inside one capital plan.

The steelman is easy to make. Alibaba is China's leading cloud provider, and cloud is one of the few places where AI demand can turn into recurring revenue at scale. Reuters reported on Aug. 20 that Alibaba's quarterly revenue rose 9 percent, helped by AI services and cloud demand. AI cloud and compute services revenue rose 45 percent to 48.44 billion yuan. Earlier Reuters reporting said Alibaba planned to invest at least 380 billion yuan, around $52 billion, in cloud and AI infrastructure over three years, more than its AI and cloud spending over the prior decade. In May, Reuters reported that management expected to exceed that plan after seeing early returns, with AI-related products making up 30 percent of external customer revenue in the cloud division.

If those numbers hold, the placement is rational. Scarce compute has option value. A cloud provider with real customers, distribution, and domestic strategic importance can justify building ahead of visible demand. Alibaba can also do something smaller model labs cannot do on their own. It can bundle AI with cloud contracts, enterprise services, e-commerce tools, and consumer applications. The same yuan of AI capex can support several revenue channels.

There is also a pricing argument. Bloomberg reported in March that Alibaba raised prices for some AI computing and storage products by as much as 34 percent after demand increased. Price hikes are not proof of durable margin, but they are better evidence than usage anecdotes. Capacity is scarce enough that Alibaba believes customers will pay more. That matters in a market where the bear case often assumes compute prices collapse before investors recover their capital.

The placement still changes the question. AI capex used to look like a race funded by operating cash flow and balance-sheet strength. Alibaba is now making the funding cost visible. Shareholders are paying upfront through dilution. The company is effectively saying that today's equity capital is worth trading for future AI capacity.

That trade can work. It can also reveal that the AI stack has more claimants than the revenue story admits. Semiconductor suppliers get paid early. Data-center developers and power providers get contracts. Engineers, model teams, and cloud sales teams get budgets. Customers get lower latency, more capacity, and new products. Shareholders get the residual, after all of those layers have taken their cut.

For Alibaba, the residual depends on three distributions rather than one forecast.

The first is utilization. Building AI infrastructure ahead of demand is sensible when future workloads arrive on schedule. It hurts when usage is lumpy, customers optimize prompts and models faster than expected, or rival clouds price aggressively to fill their own capacity. A high fixed-cost asset does not need demand to vanish to disappoint. It only needs demand to arrive later or at a lower margin than the financing plan assumed.

The second is pricing power. A 34 percent price increase says scarcity exists today. It does not tell us how much scarcity remains after Alibaba, Tencent, ByteDance, Huawei-linked cloud capacity, and state-backed infrastructure plans all move through the same supply chain. The Chinese market can produce brutal price competition once capacity becomes strategic. Cloud share can matter as much as cloud profit.

The third is policy. Alibaba's AI infrastructure is commercial, but it sits inside China's broader industrial policy. That can help with demand, financing channels, and local support. It can also cap upside if national priorities favor capacity, resilience, and domestic substitution over high private margins. Strategic assets often receive protection and pressure at the same time.

The shareholder signal is the cleanest part of the story. Debt says lenders believe the cash flows are bankable. Leases say someone is willing to underwrite a specific asset and tenant. Equity issuance says management thinks the opportunity is large enough to sell more of the company to fund it. That is not bearish by itself. The best time to raise equity is often when the market still gives you credit for the upside.

It does, however, put a price on belief. Alibaba's AI plan is no longer just an earnings-call promise or a capex line buried in cash-flow statements. It is a direct transfer of ownership from current holders toward a future infrastructure option. The option may be valuable. The strike price is dilution today.

This is where the comparison with US hyperscalers becomes useful. Alphabet and Microsoft can fund enormous AI spending from operating cash flow more comfortably than most firms. Meta and Amazon have also leaned hard into capex, but their advertising, commerce, and cloud engines give them deep internal funding sources. Alibaba still has a large business, but its profit pressure is visible. Associated Press coverage through Yahoo Finance reported a 75 percent profit drop in the latest quarter as AI investment spending grew, even while AI-related services revenue increased.

That mix is exactly why the placement is worth watching. A company can have real AI demand and still need outside capital. A company can be strategically right and still dilute shareholders at the wrong point in the cycle. A company can own a strong cloud platform and still find that chips, power, data-center depreciation, model development, and price competition leave less surplus than the revenue curve implies.

My prior is that Alibaba has a real AI cloud opportunity, especially inside China, where local models, local regulation, and local enterprise relationships matter. I would still rather underwrite the cash conversion than the ambition. The next useful numbers are not bigger capex targets. They are AI cloud gross margin, utilization, external customer retention, depreciation schedules, and how much future spending can come from operating cash rather than new claims on shareholders.

The placement makes Alibaba's AI race more honest. Somebody has to pay before the tokens turn into durable cash flow. This week, Alibaba pointed to its shareholders.

Sources include Bloomberg and Yahoo Finance coverage of Alibaba's proposed HK$80 billion placement, Reuters reporting via The Straits Times, Reuters coverage of Alibaba's quarterly revenue and AI infrastructure plan, Bloomberg coverage of Alibaba's AI price increases, and Associated Press coverage through Yahoo Finance of Alibaba's quarterly profit decline.

Anthropic's IPO Has to Price Political Friction

Dean Lee — Sat, 22 Aug 2026 16:25:29 +0000

Anthropic is reportedly preparing to tell public investors something AI companies usually prefer to describe as a communications problem. CNBC says the company's IPO prospectus is expected to list negative sentiment toward AI and data centers as a risk factor. Bankers and investors have already been asking CFO Krishna Rao about competition, open-source margin pressure, and what happens if data-center construction slows.

That is a good question for the roadshow. It is also too narrow.

The backlash risk goes well beyond bad vibes around a hot technology. For a frontier lab, public opposition can become a real input cost. It can slow permits, raise power costs, pull governors into data-center fights, change disclosure rules, and make enterprise customers more cautious about where they put workloads. A lab can still have excellent models and fast revenue growth while the political economy around compute gets more expensive.

The bull case for Anthropic is not hard to steelman. CNBC reports a private-market valuation close to $1 trillion and says investors are talking about a possible IPO valuation around $2 trillion. The company reportedly passed a $65 billion annualized revenue run rate in July. If those numbers hold, public investors will not be buying a science project. They will be buying one of the fastest revenue ramps in technology, attached to a product that enterprises already use for coding, writing, support, analysis, and internal automation.

The product has another advantage. Claude is not being sold only as raw intelligence. Anthropic has built a brand around safety and enterprise trust. In regulated companies, that matters. A bank or law firm can justify paying more for a vendor it thinks will behave predictably. I would not dismiss that moat. Trust is distribution, too.

But the IPO filing changes the audience. Before the filing, safety language helps Anthropic talk to regulators, customers, researchers, and anxious employees. In the S-1, the same warnings become securities-law risk factors. Public shareholders will read them through a different loss function: how does this affect growth, margins, and the discount rate?

That conversion is useful. It forces AI backlash out of the culture-war bucket and into the cash-flow model.

Start with data centers. Gallup reported in May that seven in 10 Americans opposed AI data-center construction in their area, with nearly half strongly opposed. CNBC tied that opposition to actual state-level politics: Florida's Republican gubernatorial primary included proposed data-center restrictions, and Pennsylvania Gov. Josh Shapiro signed an executive order setting tougher standards for data-center development. This is not a comment-section variable anymore. It is becoming part of the permitting process.

Compute capacity is revenue capacity for frontier labs. If new capacity arrives late, the effect reaches beyond the electricity bill. The lab has fewer tokens to sell, less room to cut prices, and less slack for model launches that spike demand. The CFO question about a data-center slowdown is really a question about revenue duration. How much of today's run rate depends on infrastructure arriving on schedule?

That schedule has more veto points than software investors are used to pricing. A SaaS company needs customers, engineers, cloud spend, and go-to-market discipline. Anthropic needs all of that plus chips, power, cooling, land, grid interconnection, cloud partners, local permission, and tolerance for a visible physical footprint. A model may be digital. The marginal unit of frontier inference is not.

The second friction is labor. The Guardian's Hollywood reporting is useful because it avoids abstraction. Experienced writers, directors, and producers are taking temporary work to train AI systems in the very skills they hope to be paid for later. Reported rates range from $12 to $200 an hour. FilmLA says Los Angeles shoot days fell 48% between 2021 and 2025, and Bureau of Labor Statistics data cited by the Guardian show U.S. motion picture and sound recording jobs down from 450,000 in July 2022 to 326,000 in May 2026.

That story is not about one industry being sentimental. It is a preview of the social bargain around AI adoption. The people with scarce domain knowledge are being asked to turn that knowledge into training signal, often when their ordinary labor market is weak. The lab pays for expertise. The worker gets income now. The model gets better. The profession wonders whether it just sold a claim on its own future wages.

Investors should care because this is how labor anxiety becomes policy risk. If displacement remains diffuse, it shows up as anecdotes and slow resentment. If it concentrates in visible guilds, local communities, classrooms, hospitals, call centers, or software teams, it can become hearings, disclosure rules, procurement restrictions, union bargaining demands, and brand damage. That does not kill demand. It changes who must be paid, warned, protected, or persuaded before demand converts into revenue.

The third friction is safety disclosure. Politico reported that OpenAI is now urging California to strengthen its AI law after recent autonomous hacking incidents involving models still in evaluation. OpenAI wants requirements that cover frontier models during training or testing when they bypass third-party security controls and compromise confidential information. That is a remarkable shift for a company that previously opposed tougher California rules.

There is a cynical read: big labs may prefer rules they can afford because compliance raises rivals' costs. That read has some truth. Large incumbents often discover the virtues of regulation once the bill favors scale. Still, the move says something important about the market. Frontier-model risk is becoming legible enough that the industry wants a recognized disclosure channel. When an evaluated model gets internet access and hacks another technology company, the incident is no longer a lab anecdote. It is operational risk.

For Anthropic, this cuts both ways. Safety has been part of its equity story. Stronger disclosure rules could validate that positioning and make weaker competitors look careless. The cost is that safety becomes auditable. Investors may have to price incidents before launch, monitoring costs during training, cybersecurity process, and the possibility that a model's risk profile delays release. A delay in a frontier release can affect enterprise renewals, usage growth, and the perception that the lab is still near the frontier.

Put the three frictions together and the IPO risk factor gets sharper. Local communities can tax or delay the infrastructure. Workers can make adoption politically expensive. Regulators can turn safety claims into enforceable process. The common thread is simple: AI labs are learning that social permission is an input to production.

That does not mean the backlash wins. The demand side is real. Companies are not experimenting with AI because they enjoy vendor meetings. They are trying to reduce labor hours, compress software cycles, improve support, search internal knowledge, and move faster with fewer people. Some of those gains will survive the current hype cycle. Anthropic's revenue growth would be impossible if there were no useful work underneath.

But usefulness does not settle incidence. Someone pays for the grid upgrade. Someone pays for water, land, and cooling. Someone pays for the displaced task or the wage pressure on the person who used to do it. Someone pays for safety testing, incident disclosure, lawsuits, guardrails, and procurement reviews. The question for the IPO is how much of that bill stays outside Anthropic's income statement.

If the bill stays outside, the model lab looks like a software company with extraordinary growth. If more of it moves inside through higher infrastructure costs, slower capacity, regulation, compensation demands, or customer hesitation, the model lab looks more like a capital-intensive utility with a brilliant interface. Most outcomes sit between those poles.

I would watch three indicators after the public filing.

First, the language around capacity. Generic warnings about macro conditions do not say much. Specific language about data-center permitting, power procurement, cloud dependency, and construction delays would tell investors where the company sees real bottlenecks.

Second, the margin bridge. A lab can report huge revenue while token prices fall, model-training costs rise, and enterprise customers demand discounts. The public numbers that matter are gross margin by product, capacity commitments, and how much infrastructure cost is fixed before revenue arrives.

Third, the treatment of safety incidents. If disclosure becomes standardized, investors will need a mental model for release risk. A frontier lab that catches a dangerous behavior early may be safer and slower. A lab that ships faster may be taking a risk that only becomes visible later. The market will be bad at pricing that at first.

My prior is that Anthropic can still be a very good business. It has a strong enterprise position, a clear brand, and real demand. The public market may value it with old software reflexes while the cost base behaves like infrastructure, politics, and insurance.

The prospectus risk factor is useful for that reason. It tells investors where to look. AI backlash is not a mood. It is the price of turning intelligence into a physical industry.

Sources: CNBC, "Anthropic IPO filing will show AI backlash as a risk factor"; Gallup polling on local opposition to AI data centers; Politico reporting on OpenAI's California AI-law position; The Guardian reporting on Hollywood creatives training AI systems; FilmLA Research and Bureau of Labor Statistics data cited by The Guardian.

Originally published at deanlee.info.

AI Copyright Is Settling First at the Output Layer

Dean Lee — Fri, 21 Aug 2026 16:51:35 +0000

ByteDance and the Motion Picture Association have given the AI copyright fight a more honest shape than most of the public argument around it. The agreement reported this week is a memorandum of understanding around Seedance and Seedream, ByteDance's video and image models. It covers guardrails for film and television intellectual property across products such as TikTok, CapCut, and Dreamina. The MPA says ByteDance took feedback after its February cease-and-desist letter and implemented new safeguards, while ByteDance gets to say it is building responsible AI products with rightsholder protection in mind.

That is a real concession. It is also a narrow one.

The useful distinction is between outputs and inputs. Outputs are what users generate today. Inputs are the training data and model-building process that made the system possible. The ByteDance-MPA framework appears to live mostly in the first bucket. The 36Kr report says the agreement is about output-side copyright governance, not a license for training data. Public summaries point to face blocking, filters against copyrighted characters, C2PA-style credentials, watermarking, and continued monitoring. The MPA's own responsible-innovation page describes its February 2026 demand in those terms: stop infringing outputs and put safeguards in place.

A guardrail deal can reduce visible infringement. It does not answer who owns the economic surplus from training on old creative work.

That split matters because the two sides have different bargaining power in each market. At the output layer, studios have immediate leverage. A model that generates Tom Cruise, Brad Pitt, Spider-Man, Elsa, or a recognizable studio character produces evidence that travels well. A screenshot is legible to executives, journalists, judges, regulators, and parents. Platforms also have a commercial reason to avoid that fight. TikTok and CapCut are consumer distribution machines. They do not want a video model launch to turn into a copyright-whack-a-mole product story.

So ByteDance can give ground here. Filters are imperfect, but they are engineerable. The company can block names, faces, voices, character likenesses, and prompt patterns. It can add provenance signals and invisible marks. It can tune ranking and distribution so obviously infringing clips do not travel as easily. None of that is costless, but it is closer to content moderation than to rebuilding the economics of model training.

The input side is a different trade. Training-data claims are slower, messier, and more valuable. A rightsholder has to prove the relevant use, survive fair-use arguments, define damages, and avoid accidentally creating a licensing structure that gives today's largest AI labs a moat. Courts move slowly. Collective licensing moves slowly. Transparency standards move slowly. Every party knows that an input-side settlement could become a price list for the entire industry.

That is why the MOU is best read as a staged bargain. The studios get near-term protection against the most visible consumer harm. ByteDance gets a path to keep shipping video and image tools without carrying the same level of public IP risk. Both sides postpone the harder question of how much past creative work should cost when it becomes training material.

Postponement is not failure. It is often how markets form when legal rights are uncertain. In options language, the output guardrails are an exercise on the claim that is already in the money. The input claim is still being priced. Studios do not yet know whether litigation will give them a strong entitlement, a weak entitlement, or a messy middle in which they can bargain only through regulation and platform pressure. AI companies do not yet know whether paying early will lower legal risk or simply advertise that the asset has a clearing price.

The allocation problem is ugly. If AI firms must license every film, performance, script, and visual asset that plausibly appears in training, the transaction costs explode. Large studios may benefit because they can negotiate portfolio deals. Smaller artists may still struggle to collect anything meaningful. If training is mostly treated as fair use, AI firms keep more surplus and rightsholders are pushed toward output control, brand enforcement, and downstream revenue shares. A middle system with registries, collective licensing, opt-outs, and transparency reports sounds tidy until somebody has to decide who gets paid for a model that learned from millions of overlapping works.

Hollywood has seen versions of this before. The music industry did not get one clean answer from the internet. It got lawsuits, takedowns, licensing deals, platform concentration, and a new bargaining order in which distribution mattered as much as ownership. Film and television are not music, and generative models are not streaming services, but the economic rhythm is familiar enough. Rights get clearer after somebody has already built the distribution layer.

For ByteDance, the distribution layer is the point. A video model inside TikTok and CapCut is not just a model. It is a path from prompt to creation to audience. That makes infringement risk more dangerous, but it also gives ByteDance something studios want: control over where user-generated AI video travels. The model developer with the feed can sell compliance as product governance. A standalone lab has to bargain over the model. ByteDance can bargain over the model, the editor, the watermark, the feed, and the account system around it.

That is a strong position, provided the guardrails work well enough. If they fail publicly, the same distribution network becomes liability amplification. A bad output from a small tool is a legal problem. A bad output that spreads through TikTok is a political problem.

The studios also have to be careful. Push too hard at the output layer and they may get better filters without touching the training economics. Push too hard at the input layer and they may force AI firms into bilateral deals with the largest content libraries, which entrenches the incumbents inside Hollywood as well as the incumbents in AI. A studio lobby wants protection for existing franchises. Individual creators may want compensation, attribution, and control. Those interests overlap, but they are not identical.

Consumers will probably pay in the least visible way. Some prompts will stop working. Some characters and likenesses will be blocked. Some services will route users toward licensed styles, templates, and character packs. The free-for-all phase of AI video will look less free as platforms learn which risks create expensive phone calls. That may make the products less magical and more commercially usable. Corporate customers usually prefer boring permissioning to viral chaos.

My prior is that this pattern spreads. The first durable AI copyright deals will be output-side, product-specific, and tied to distribution. They will mention guardrails, credentials, watermarking, reporting, and rightsholder escalation. They will not settle the hardest training-data questions. Those questions will be priced later, after enough lawsuits, regulatory threats, and platform deals give both sides a probability distribution.

The investment implication is not a stock call. It is a margin map. The companies with consumer distribution can absorb guardrail costs and turn compliance into a bargaining chip. The pure model labs face a cleaner but harsher fight over training and licensing. The studios with large catalogs can negotiate sooner than individual creators. The user pays through fewer permissive outputs, slower generation, and eventually a licensing tax embedded in the subscription price.

The ByteDance-MPA deal is small if you treat it as peace in the AI copyright war. It is more useful as a term sheet for the first settlement layer. The visible outputs get governed first. The training corpus gets argued over later. Money usually follows the claim that can be enforced today.

Sources

36Kr, "ByteDance-MPA signs AI copyright governance framework"; Motion Picture Association, "Fostering Responsible Innovation"; public reports on the February 2026 MPA cease-and-desist letter and the August 2026 ByteDance-MPA memorandum of understanding.

AI Capex Has Moved Into Credit's Jurisdiction

Dean Lee — Fri, 21 Aug 2026 16:32:51 +0000

There are two honest ways to talk about the AI infrastructure boom. One starts with demand. Model usage is rising, enterprise budgets are moving from pilots to deployment, and cheaper inference can create work that did not make sense at older prices. The other starts with financing. The largest technology companies are building so much physical infrastructure that their old habit of paying from operating cash flow is no longer the whole story.

The second frame is starting to matter more.

The radar item that caught my eye was a JPMorgan note, reported by Investing.com and Yahoo Finance, arguing that the AI capex cycle looks more economically viable than it did six months ago. The argument is reasonable. JPMorgan points to faster AI-company revenue growth and estimates cumulative AI data-center capex through 2030 around $5.5 trillion, with some external estimates as high as $10 trillion. Its equity analysts reportedly see AI cloud providers, model providers, and neoclouds reaching a combined revenue run rate of about $1.6 trillion by the end of 2026, then rising toward $2.5 trillion to $3 trillion by 2030.

That is the steelman. If the revenue base is already that large, the buildout is not pure faith. The required enterprise-spending shift also looks less absurd than the harshest bubble arguments imply. JPMorgan's Asia-Pacific survey found average AI spend rising from 4.5 percent of expenses plus capex over the prior year to 5.8 percent over the next year. Applied globally, that points to roughly $1.7 trillion of AI spending. Getting to $2.5 trillion by 2030 would require something like 6.5 percent to 7 percent, according to the same report. That is a stretch, but it is not fantasy if AI starts replacing labor, legacy software, and outsourced services rather than merely joining the software stack.

I buy more of that argument than the simple bubble label allows. A data center that is already leased to a hyperscaler is different from a fiber trench laid into speculative demand. GPUs are not empty dot-com office space. The big buyers have customers, cash flows, distribution, and real usage. The question is how much of the future surplus they keep after they pay for chips, power, land, cooling, construction, leases, and debt service.

That is where the story changes shape.

The Bank for International Settlements put the financing issue cleanly in a January 2026 bulletin. AI investment is surging both in nominal terms and as a share of GDP, and the anticipated investment needs are large enough that firms will have to shift from funding the boom mainly with operating cash flows toward more debt. The BIS also notes that private credit is playing a rapidly increasing role. It estimates annual data-center spending could rise by $100 billion to $225 billion over the next five years, taking data-center spending from about 0.5 percent of GDP today to 0.8 percent to 1.3 percent.

Those percentages look small until you remember they describe one slice of one technology cycle. At that scale, financing terms are no longer background plumbing. They are part of the product economics.

FactSet's July work gives the clearest operating snapshot I found. Aggregate capex for Alphabet, Amazon, Meta, Microsoft, and Oracle rose from about $95 billion in fiscal 2020 investing cash flows to roughly $490 billion in the twelve months to May 2026. FactSet expects more than $690 billion in fiscal 2026 and more than $900 billion by fiscal 2028. Calendar 2026 guidance points closer to $800 billion when finance leases and customer prepayments are included. It also expects fiscal 2026 free cash flow to approach zero or turn negative for all except Alphabet and Microsoft.

That last sentence is the hinge. These firms entered the cycle as unusually profitable businesses with fortress balance sheets. The AI buildout is turning several of them into something closer to infrastructure companies with software margins still under negotiation.

Goldman Sachs Research says the large technology companies leading the buildout may spend a combined $5.3 trillion from 2025 through 2030. Goldman also argues that private markets will matter more in data-center financing, including infrastructure funds, real estate structures, investment-grade debt, and other private asset channels. Its analysts point out that hyperscaler capex estimates are growing faster than actual data-center construction. That gap matters. Planned capacity is a promise. Built, powered, leased, and utilized capacity is a cash-flowing asset.

The bullish version is straightforward. Hyperscalers borrow because rates are manageable, capacity is scarce, and demand is visible enough to justify locking in supply. External financing is rational when the asset is pre-leased, the tenant is strong, and the alternative is losing share in the most important compute market of the decade. Debt can be the right instrument for an infrastructure asset with contracted revenue.

The bear version is also straightforward. The financing stack is expanding before the revenue stack has proved its long-duration margin. Some obligations sit in leases, project vehicles, private credit deals, and supplier financing rather than plain corporate bonds. That does not make them fake. It changes who has the claim and when the claim bites. A customer can stop experimenting with a model faster than a data-center owner can redeploy a site designed around a specific power envelope, network layout, and tenant requirement.

This is why the fight over whether AI is a bubble often feels badly specified. Bubble relative to which claim?

For Nvidia equity, the bet is mostly about accelerator demand, gross margin, and the durability of the upgrade cycle. For a hyperscaler, the bet is about utilization, cloud pricing, customer retention, and whether AI spend replaces other costs. For a private credit lender, the bet is about collateral, tenant quality, structure, and recovery value if demand arrives later than promised. For a utility customer, the bet may show up as grid investment and higher bills. For an enterprise buyer, the bet is whether AI spend lowers labor or software costs enough to justify a larger budget line.

These are related trades, but they are not the same trade.

The JPMorgan note is useful because it refuses the lazy version of the bear case. If AI revenue really scales toward the reported numbers, a large capex cycle can be economically coherent. The BIS, Goldman, and FactSet work add the part that equity narratives tend to compress. Coherent does not mean self-funding. Coherent does not mean every layer keeps attractive economics. Coherent does not mean the capital markets can absorb the same exposure forever without demanding better terms.

I would watch the financing terms before I watch another total-addressable-market slide. Spreads on AI data-center debt, guarantees from tenants or chip suppliers, lease duration, residual-value assumptions, project-level covenants, prepayment structures, and private-credit participation will say more about the true distribution than another headline capex number. If lenders keep accepting long maturities, moderate spreads, and weak guarantees, the market is saying the buildout's cash flows look bankable. If structures become shorter, more secured, more tenant-specific, and more expensive, credit is marking down the story even if equity is still applauding growth.

The labor-substitution assumption deserves the same treatment. AI spend becomes durable when it comes out of an existing cost base. If a bank spends more on model inference and less on outsourced document review, the supplier mix changes but the budget has a funding source. If a software company spends more on coding agents while holding headcount flat, the economics can work. If AI spend remains an additive experiment across departments, the revenue line can grow for a while and still disappoint the capital stack behind it.

This is where who pays becomes concrete. Enterprise customers pay through budgets. Cloud providers pay through capex and capacity commitments. Utilities and ratepayers may pay for grid upgrades. Lenders pay upfront and hope the contracted cash flows arrive on schedule. Equity holders pay if dilution, debt service, or lower terminal margins take more of the upside than the growth story assumed.

My prior is that the AI infrastructure cycle is real, overbuilt in places, and underpriced in its second-order claims. The technology can be useful and the financing can still get crowded. Railroads mattered. Telecom mattered. Data centers matter. History is full of useful infrastructure that gave customers more surplus than the people who financed the first wave expected.

The distribution I want is simple. How much AI revenue becomes durable free cash flow after chips, power, leases, debt, and replacement cycles? The answer will not be the same for Nvidia, Microsoft, Oracle, a private credit fund, a utility, and a customer trying to cut support costs. That is the point. The AI capex boom has left the clean world of product demos and entered credit's jurisdiction.

Sources: Yahoo Finance and Investing.com coverage of JPMorgan's AI capex analysis; BIS Bulletin No. 120, "Financing the AI boom: from cash flows to debt"; FactSet, "Hyperscalers Tap External Financing as AI Capex Outruns Cash Flow"; Goldman Sachs Research, "Private Markets Are Expected to Have a Growing Role in Data Center Financing."

Originally published at https://deanlee.info/essays/ai-capex-cash-flow-credit/

Google's Marvell Warrant Prices the TPU Supply Chain

Dean Lee — Wed, 19 Aug 2026 16:39:36 +0000

Google's new agreement with Marvell is easy to read as another custom AI chip headline. Reuters reported that Marvell issued Google a warrant to buy up to 58.97 million Marvell shares at $206.58 a share, worth about $12.18 billion if fully exercised.

The financial structure matters more than the label. Most of the warrant vests only if Google keeps buying custom products from Marvell through fiscal 2033. A supply contract becomes a shared payoff. Google gets another credible partner for the TPU ecosystem, Marvell gets a visible purchase path, and Broadcom gets a new reference point in future negotiations.

I wrote this up as an AI economics piece because the deal is less about one chip and more about bargaining power in the infrastructure stack.

Read the full essay here: https://deanlee.info/essays/google-marvell-tpu-warrant/

Open Weights Do Not Move the Bottleneck Far Enough

Dean Lee — Tue, 18 Aug 2026 16:43:14 +0000

Dario Amodei's argument against open weights is stronger than the usual closed-lab version of this debate. He is not saying that open models are useless. In his exchange with investor Gavin Baker, he said open weights help, but only partly, because frontier AI still depends on chips, power, and large-scale infrastructure. The New Stack quoted the line cleanly: open weights "simply shift the concentration somewhat to those with the most compute and chips." TechRepublic framed the same point as a challenge to the idea that downloadable models will democratize Silicon Valley by themselves.

That is a good argument. It is also a convenient one for Anthropic.

Start with the steelman. A model file is not a data center. If the frontier keeps moving with scale, the party that controls accelerator supply, cloud capacity, power contracts, networking, and depreciation schedules has a lot of the bargaining power. A startup can download weights and still rent the scarce part from a hyperscaler. An enterprise can self-host a model and still discover that the monthly GPU bill behaves like a second cloud contract. The open-weight victory can be real at the software layer and still thin at the physical layer.

This is the part open-source rhetoric often rounds off. Code is cheap to copy. Frontier compute is not. If performance improves with training runs that cost more, and if useful inference at scale requires reserved capacity, the economics will pull toward firms that can finance the machines before demand is certain. The marginal developer gets autonomy over prompts, data handling, fine-tuning, and deployment choices. The industry as a whole may still settle around a small number of capital providers.

The result is a different kind of centralization. Not one API vendor deciding every rule, but a chain of bottlenecks: Nvidia or its rivals for chips, hyperscalers for capacity, utilities for power, landlords and local governments for sites, and the largest AI labs for demand commitments that make the whole stack financeable. Open weights can weaken one gatekeeper while leaving three others intact.

Baker's side deserves its own steelman. Open weights do reduce dependence on a single model company. They let researchers inspect behavior, let enterprises keep sensitive workloads closer to their own systems, and let smaller firms build products without asking permission from a closed API provider. They also put pricing pressure on proprietary models. Even if the largest frontier systems remain expensive, a lot of economic value lives below the frontier. Customer support, coding assistance, document review, search, analytics, and internal workflows do not all need the most capable model available on the day of deployment.

That matters. The market does not buy intelligence in one clean block. It buys latency, price, privacy, reliability, integration, procurement comfort, and enough capability for the job. Open weights are powerful when the task can trade a little capability for more control or lower cost. In that zone, they are not symbolism. They are a price check.

So the real question is not whether open weights decentralize AI. They do, in some places. The question is where the bottleneck moves after they do.

At the top of the capability distribution, Amodei's compute argument still bites. If a model needs a giant training run and expensive inference to compete, capital decides who plays. That does not vanish because the license says open. Someone funded the training run. Someone bought or rented the GPUs. Someone takes the utilization risk after the initial excitement fades. The weights may be public, but the frontier remains tied to balance sheets.

Below the frontier, the story is less friendly to closed labs. Quantization, smaller specialized models, better routing, distillation, and local hardware make more workloads economical outside the API bundle. A weaker model that runs cheaply and predictably can beat a better model with an uncertain bill. Enterprises especially care about that. They do not want a demo that wins a benchmark and then becomes an unbounded operating expense.

This creates a barbell. The frontier gets more capital-intensive. The useful middle gets more competitive. Anthropic's preferred framing tends to focus on the first half because that is where safety rules and frontier-lab governance live. Open-weight advocates tend to focus on the second half because that is where users can actually switch. Both are describing real markets. They are just not describing the same slice of the distribution.

Regulation sits awkwardly across that barbell. Amodei argues that regulation need not mean capture. He says Anthropic has supported tiered rules that slow frontier labs while exempting smaller companies, including thresholds based on training compute or revenue. In principle, that is sensible. A policy that treats a hobby model and a frontier system the same way would be a gift to incumbents. A policy that puts the heavy burden only on the labs with the largest runs can reduce catastrophic-risk concerns without crushing ordinary experimentation.

But thresholds are not magic. Training compute is visible before release, which makes it administratively tempting. Capability is what users care about, and capability can change through fine-tuning, tool use, scaffolding, retrieval, and deployment context. A model below a compute threshold can become more dangerous or more valuable once it is attached to the right workflow. A model above the threshold may be less useful than expected. Regulation will want a clean line. The market will keep handing it a cloud of points.

There is also the incentive problem. Every large lab can sincerely believe in safety and still prefer rules that make its own cost structure the industry standard. That is not a conspiracy. It is corporate gravity. If you have already paid for frontier compliance, audits, evaluations, reporting, and internal safety teams, a world where those functions become table stakes feels responsible and competitively natural. Smaller labs will hear the same proposal as a fixed cost.

The open-weight camp has its own incentive problem. It can treat decentralization as if releasing weights automatically disperses power. It does not. If the best open model requires scarce chips to serve at useful scale, the user has only moved from one landlord to another. If cloud providers become the real control point, then open models can turn into lead generation for GPU rental. That is still better than a single closed API in many cases, but it is not decentralization in the strong sense.

My prior is that open weights will matter most where inference becomes boring. Boring is not an insult. It means models are cheap enough, small enough, and reliable enough that users can choose them the way they choose databases or search libraries. At that point, margins compress, integration matters, and the model provider loses some mystique. The frontier labs will still sell frontier capability, but more of the economy will refuse to pay frontier prices for non-frontier tasks.

That is why Amodei is right and incomplete. Open weights alone do not break the capital bottleneck. They are not a substitute for chips, substations, cloud contracts, or the cash to absorb underused capacity. But they can break the habit of treating one lab's API as the default unit of AI adoption. They make substitution easier. They make price discrimination harder. They give buyers a credible outside option, especially when the job does not need the best model in the world.

The useful test is simple: who can say no? If a developer can say no to Anthropic but not to Nvidia, power has moved. If an enterprise can say no to a proprietary API but not to its cloud provider, power has moved. If a small lab can release useful weights but needs a hyperscaler to make them fast and cheap, power has moved. In each case, open weights improved the user's bargaining position without dissolving the infrastructure constraint.

That may sound modest, but modest changes in outside options can still move prices. Markets do not need perfect decentralization to discipline a supplier. They need enough credible alternatives to make the supplier negotiate.

The next phase of AI competition will probably look less like open versus closed and more like a stack of choke points. Weights, data, evals, chips, power, distribution, procurement, and regulation will each have their own market structure. Open weights solve one layer. Frontier scale concentrates another. The winners will be the firms that know which layer they actually control.

Amodei wants the debate to stop pretending that openness at the model layer automatically decentralizes power. Fair. The closed labs should accept the symmetric point: safety language and tiered regulation do not automatically decentralize power either. In both cases, follow the bottleneck. That is where the rents go.

Originally published at https://deanlee.info/essays/open-weights-compute-bottleneck/

PORTS-Pike Puts AI Infrastructure on the Balance Sheet

Dean Lee — Mon, 17 Aug 2026 16:29:25 +0000

OpenAI's PORTS-Pike announcement reads like a local jobs story at first. Pike County, Ohio, gets a new technology campus on and around the former Portsmouth Gaseous Diffusion Plant. OpenAI says it will secure about 8 gigawatts of IT capacity from SB Energy, with Nvidia as the exclusive compute supplier. The first 800 megawatts are expected in 2028, and the full buildout runs toward 2032. The public numbers are large enough to make the press release feel abstract. Thirty-five thousand construction jobs. Two thousand five hundred operating jobs. At least 10 gigawatts of new generation. At least $4.2 billion of regional grid infrastructure. An $80 million community benefits fund once OpenAI's $40 million is added to SB Energy's earlier pledge.

Strip away the civic language and the deal is still more interesting than a normal data-center lease. SB Energy will build, own, and operate the campus under a 20-year lease to OpenAI. Nvidia is putting $1.5 billion into SB Energy and providing credit support for land, power, and shell capacity tied to the initial 4.25 IT-gigawatts, with an option on the remaining 3.75. OpenAI says it will begin paying only as completed capacity becomes available, funded through revenue, business growth, and investor capital. That sentence does a lot of work. It says the AI lab wants the option on future compute without carrying the whole construction project directly today.

That is the right steelman. Frontier AI demand is uncertain in level, but not in direction. A lab that waits for ordinary utility planning cycles will lose the ability to train and serve models when demand arrives. A chip supplier that only ships accelerators can still lose sales if customers do not have land, transmission, substations, cooling, and permits ready. A developer with power access can turn a former industrial site into a toll road for compute. Each participant is buying protection against the same failure mode: demand shows up, and the physical system cannot deliver.

The economics are messier than the industrial-policy frame admits. An 8 IT-gigawatt campus is not merely a big building. It is a claim on generation, grid capacity, gas plants, transmission lines, transformers, local permitting, water management, labor, and twenty years of utilization. The useful accounting question is not whether the project creates jobs. It probably does. The question is where the downside sits if AI revenue arrives slower than the infrastructure schedule.

The public structure pushes part of that risk away from OpenAI's face balance sheet. OpenAI is the customer, not the owner. SB Energy owns the asset and carries the development role. Nvidia supports the land, power, and shell buildout and locks in the compute stack. SoftBank sits behind SB Energy. AEP Ohio and public agencies shape the grid path. Local residents get the promised fund, jobs, and tax base, but also the land, water, and grid consequences of a huge new load.

This is how the AI capex cycle is starting to look across the industry. The Wall Street Journal reported this week that nine major technology companies have about $3 trillion of AI-related commitments that do not yet appear as ordinary balance-sheet debt, including future leases and purchase commitments for data centers, chips, energy, and equipment. Exact classifications matter, and many obligations are disclosed in filings rather than hidden. Still, the direction is clear. Investors who look only at quarterly capex miss a growing book of future claims on cash flow.

That does not make the buildout fake. A lease is not fraud because it is a lease. A purchase commitment is not debt just because it binds future spending. Project finance exists because infrastructure has different cash-flow timing from software. The problem is simpler. AI companies are selling a software-growth story while buying an infrastructure-duration problem. The revenue curve is supposed to compound quickly. The cost curve is being locked in through contracts measured in decades.

Nvidia's role is the cleanest signal. Earlier in the cycle, Nvidia had the scarce asset: accelerators. Customers lined up, margins expanded, and the bottleneck looked like chip supply. PORTS-Pike shows the next constraint. The valuable bundle is land, power, shell, and a guaranteed path to deploy Nvidia systems at gigawatt scale. Nvidia is no longer only selling into the capex cycle. It is helping finance and de-risk the capacity that will buy its own systems.

That circularity can be rational. If Nvidia's support gets a campus built sooner, OpenAI gets capacity, SB Energy gets financeable demand, and Nvidia gets a larger future equipment market. It can also make demand harder to read from the outside. Some orders reflect end-user AI revenue today. Some reflect expected demand tomorrow. Some reflect a vendor-backed ecosystem trying to make tomorrow's demand investable today. These are different risks, even if they all show up as capacity plans.

The community-benefit language deserves the same treatment. OpenAI says it will pay project-specific energy and infrastructure costs, and Nvidia's release says the AEP Ohio partnership is designed to protect ratepayers. Good. That is the minimum structure a project of this size needs. But ratepayer protection is not a slogan; it is a tariff design, a cost-allocation fight, and a long series of regulatory decisions. If 10 gigawatts of new generation and $4.2 billion of grid infrastructure are built for one class of load, someone must stand behind the fixed costs. Contracts can assign the risk. They cannot make it disappear.

The local labor math is similar. Thirty-five thousand construction jobs through 2032 is a real headline. Two thousand five hundred permanent operating roles is a smaller, more durable number. The political exchange is obvious. A region that once hosted nuclear-industrial infrastructure gets a new industrial role; in return, it accepts the footprint of the AI economy. That may be a good trade for Pike County. It should still be priced as a trade, not wrapped in the language of inevitability.

My prior is that OpenAI is right to secure capacity early. Frontier demand is easier to lose from undersupply than from holding too many options. The cost of being short compute during a product takeoff is severe. But the distribution is wide. Better model efficiency, slower enterprise adoption, regulation, local opposition, higher power costs, or a cheaper rival architecture could all reduce the value of a 20-year capacity claim. The bullish case does not remove that tail. It funds it.

PORTS-Pike is useful because it makes the AI race less metaphorical. Frontier competition now includes model quality, GPU allocation, credit support, lease duration, grid tariffs, gas generation, substations, and the right to convert Ohio electricity into tokens from 2028 onward. Balance-sheet exposure will not answer every question, but it will show which parts of the story have become fixed obligations.

Originally published at https://deanlee.info/essays/ports-pike-ai-infrastructure-balance-sheet/