The same H100 costs 3x at AWS, and a four-month-old lab is catching genes the top sequencers missed

Nathan Labenz opens the weekly AI:AM highlights by saying real things are starting to happen. Jeremie and Ed Harris of Gladstone AI remain hard-boiled about U.S.–China cooperation, but Labenz heard a slight thaw: Ed describing the kind of transparency that lets the other side see you are not doing the thing they fear most, and noting it may not even be that costly, because "the Chinese are already all up in our systems." Steve Ho, head of research at Silicon Data, then names the compute gap in one sentence. Hyperscalers charge regularly, consistently, at least two to three times, sometimes more, than a typical new cloud, and they sell the difference as a differentiated product.

The rest of the montage sketches the other two blocks. Cohost Prakash, fresh off OpenAI Dev Day, argues that once a model can already build a game, the next question is whether it can build one with you in the moment and keep you in flow; latency at the same intelligence is about to matter a lot. Joel Borgan co-wrote a novel with frontier models, sketched the architecture himself, and still had to scaffold and edit or the prose came back unreadable. Daniel McKinnon of Gammo Labs is diagnosing children whose cases had gone unresolved, and his tools exist in part because a state-of-the-art model once renamed a gene mid-answer and got the case wrong. The cloned-voice introductions done, Labenz asks for feedback and starts Part One.

Ed Harris starts where diplomats always start: talking is better than not talking. Out of the early Trump–Xi contact, one possibly useful product is some kind of theoretical line between the two governments on AI incidents, a red phone in all but name. Gladstone's report is a long newsletter informed by about a dozen State Department diplomats who have sat across from China. They have tried the red-phone thing in the nuclear era. By and large, Ed says, they do not always answer. In critical phases the line is often exercised as leverage, something to take away, rather than as a collaborative project that makes everyone safer. If the CCP structurally took AI seriously, the engagement could look different. Based on the people they interviewed, the recommendation is to be realistic and keep a backup plan.

Labenz asks what the United States could demand, and verify, after a serious incident with the capabilities that exist today. Jeremie Harris refuses a detailed executable answer; nothing in that sentence, he says, has been red-teamed with the intelligence community. The caricature he will offer is thermal. Data centers throw off a hell of an energy footprint, they are huge, they have not been built to be hidden, and they take a long time to build, though he thinks hiding them could get faster. In a panic, first-order conversation, you are looking at clusters of a certain size. Prakash asks whether a cluster above that size would be a red line. Ed's version is slightly different: you cannot ask them to tear the building down, but you can ask to see the heat signatures go away. Turn off any running cluster above a threshold. GPU depreciation is the major OpEx cost, so that is a tremendously expensive ask in either direction. Jeremie calls it an insane ask. That, as of right now, is where they are.

Labenz restates the trap. The relationship is not great, capabilities keep moving, something crazy is expected, and when it arrives the default demand is outlandish because nobody invested in a lower-cost trust-but-verify environment. What should be built now? Ed: better verification, and better offensive options to ensure compliance when verification comes back dirty. Even a technique 50% better than watching gigawatts from space, he says, inventing the number, could save billions and keep the industry making money through the crisis. A community of verification startups already exists. Jeremie adds the bottleneck those companies have not priced in. When a crisis hits and a founder says they have the thing, the intelligence community still has to vet it into national technical means. Historically that takes years, even if money is no object, and both countries have to handshake. Gladstone's current work is introducing verification companies to IC elements early, putting founders on speed dial, and telling teams with $10 million-plus research agendas when an appropriately placed person would call the trajectory a nonstarter, before they burn the money.

After the sponsor break, Labenz stays with the brothers on work that could start among the labs themselves. The United States and China are, for now, on the same tech tree, meeting the same surprises, which he reads as fortunate. If anything good sits inside the OpenAI–Anthropic rivalry, it might be a test bed: a president putting those two in a room and telling them to police each other, then scaling the method across the Pacific. Jeremie does not call the idea insane. Poaching of personnel already mirrors some of the access China has at the frontier labs. Ed notes a basis of trust between two American companies that would not exist trans-Pacific. Jeremie adds that the hardware lottery is doing useful work. In track-two dialogues, Chinese counterparts consistently hear American safety talk as an attempt to curtail their development because they see themselves as behind, and therefore justified in doing things they would not grant America. Being in second place also means seeing warning shots at lower resolution. Quantum machine learning, if you wait long enough, is not relevant on the timescales that matter. The industry is doubling down on transformer MoEs with bells and whistles. That sameness, in principle, would let the two sides inspect each other's safety stacks the way the United States sometimes shared safety technology with Russia in the Cold War, though China has pushed back on that language before.

Labenz does not want to be naive, but he wants to weight positive signals. Xi's speech at the World AI Conference was friendly: credit to the United States for inventing AI, no call for an international arms race, a warning against overstretching the national security concept. What would actually update them? Ed will take conciliatory talk as better than hostile talk, then discount it as rhetoric used to shape the battle space. Unfakable signs would look different. DeepSeek and Zhipu coming to a pacing agreement because something crazy happened, their Hugging Face incident. The commission with jurisdiction over whether models can be released putting the brakes on a launch because a model that was properly ideological in testing stopped being ideological in the wild. Genuine indications that they are on the receiving end of incidents at the same level of detail as American frontier labs. That, he says, is when everyone starts worrying about the giant thing growing in the labs rather than about each other.

Jeremie offers a procedural test. Gladstone catalogued about five historical traps in U.S.–China diplomacy, a brief list of how China behaves, in the diplomats' assessment, when they are full of it. One dissenting diplomat thought some of the language objections were sincere, which is itself the epistemic problem. Flip each red flag and you get a green flag: no arbitrary fights over wording, engagement from empowered people, and qualitative progress that actually looks like progress. Xi showing up is not a red flag, though it has to be calibrated. The contrast with the past is so poor that merely looking as if you are moving can be real signal. On the American side, Labenz asks for low-cost credible steps that would show Washington is not chasing a decisive advantage, a Pax Robotica he worries the United States might actually try. Ed's cheapest stabilizer is transparency of the "we are not doing the thing you fear" variety, which may give away little that China does not already have, but should not be done unilaterally. Goodwill now looks like more academic-to-academic and startup-to-startup contact in the verification communities, which is already happening. The more people on each side who can tell their own leadership "I have spoken to them, they are not evil," the more the extremes get moderated.

Part Two is the economy of intelligence. Steve Ho's GPU rental index mixes quote prices and transaction prices, then uses machine learning to normalize contracts so that an H100 in one region can be compared with an H100 in another. An apartment index would never trust a $2,000 listing on the front of a building, because you cannot click an API and take occupancy. GPU rental often is Amazon-like: executable, so quotes belong in the sample, then get checked against trades. If a merchant just rented three nodes at $2.75, Ho will not assume a fourth is still $2.75. The next one might be gone, or $4, or $6, or $1, depending on what everyone else is quoting. The aim is broad coverage of the market as it actually is.

Labenz has been browsing the site and cannot miss the gap between neo-clouds and hyperscalers: two to four times, not a rounding error. Why does it survive, and can it be arbitraged? Ho reaches for a single-patty burger index. A street stand and a high-end steakhouse both sell burgers; after you adjust for the number of patties, some of the leftover premium is bundling you cannot honestly strip. Hyperscalers charge regularly, consistently, at least two to three times, sometimes more, than a typical new cloud. The reasons are a long legacy of other products on the platform, software analytics, safety, compliance, and enterprise customers who have been onboard for years and do not move. If the merchant blends the fries and the milkshake into the patty, declares a brand-new product, and charges three times, you stop pretending it is the same burger and put it in another category. That, Ho says, is how they get closer to a fungible unit of compute.

Prakash asks about SiliconMark, the physical benchmarking service that visits GPUs at the UUID level for health, throughput, and TFLOPS. Those numbers do not enter the price index. Ho sees GPU-to-GPU variance, the familiar lottery, but a cluster averages it out and the law of large numbers kicks in. The model currently uses six features, including geolocation, CPU, memory, term, and provider. Physical spec is not one of them. If inference someday makes compute more interchangeable and physical delivery enters the market, it might. For now, it does not.

The proprietary LLM index is up 11.5% over the last seven days, which seems to contradict the story that quality-adjusted prices are crashing. Ho says the indices have been misunderstood, partly because of the name. It is an expenditure-weighted price index normalized to a million, so usage mix can move it even when list prices do not. From late June through the middle of the month, the frontier proprietary series went from about $4 to about $1.6, more than a 50% drop, as labs released cheaper variants and other proprietary models, Meta and Grok among them, priced aggressively. The recent bounce, he suspects without having checked every ledger, is people leaning back into Anthropic's more expensive power model. Mercedes sells cheap variants and expensive ones; which cars actually move changes the average. Token prices are falling and usage is migrating toward the dear stuff. The two forces run opposite directions.

From measuring compute, the hosts turn to using it. Labenz was unmoved by Dots, the personal agents OpenAI led with at Dev Day. He has spent nine months evolving his own stack and expects to keep it. The product felt consumer, not developer, a first-party offering for ChatGPT users. His mother, he says, should just use it: no local database, no troubleshooting, a polished abstraction for people who do not want a bunch of new tricks. Prakash's first impressions of the models are sharper. GPT-6.1 Soul is the same price as GPT-6 Soul, or slightly lower, and is being sold as Astra performance at Soul price; he has used it and it is better, two or three weeks after the previous Soul. Cadence is stepping up. GPT-6 Astra Ultrafast is 8x; the old fast mode was 2x. The demo is interactive software, games assembling while you type. Latency at the same intelligence, once capability hurdles are cleared, is what he thinks comes next.

The feature Labenz actually liked is Sign in with ChatGPT: eligible users bring the subscription to participating apps. Classic try-for-seven-days tactics have gotten expensive when every new user costs tokens, especially if onboarding includes profile processing. At Waymark, the video company he cofounded, the first step is building a profile of a small business so later videos have something to draw on. That step has come down in price; at one point it was about a dollar per customer, which starts to matter in all-in acquisition math. Letting the ChatGPT plan travel is good for OpenAI, for users, and for SaaS companies, and a coding agent can implement it. He reads it as one way the lab can follow through on not eating the world.

Joel Borgan is the author of *The Receipt Horizon*, a novel set after a large AI war, written in collaboration with frontier models. He might have written it eventually; with kids and a part-time job, the activation energy was the constraint. He started well over a year ago, when context windows kept dropping earlier chapters out of the thread. He knew what he wanted made. He sketched it himself. The first half of the book was pretty well set before he engaged the models. Models are good at certain things and getting better at everything, but for prose, even with an exact chapter plan, you often get something unreadable unless you know how to scaffold it and then edit it yourself. A follow-up book is already underway, and the process has changed as the models have gotten more powerful. The collaboration made the novel better, he says, and it also left weaknesses he did not entirely account for. He knows AI writing will be controversial for a lot of readers.

Utopias want to become dystopias on the page, because without conflict the story is dull. After the war, humans are locked into what they can do; the book shows flourishing and major downsides. He did not want any character to be a mouthpiece. Things he wants created in the world are in the story, and he tried to steel-man the people who would hate them. Scaffolding meant a large-scale architecture and a lot of beats, then going back and forth with the flagship Claude and the Pro model from GPT-5 on, which did the detailed work. He works mostly in the chat interface. Context length was the huge unlock: you could finally give them the entire book. He hired professionals for copy editing, art, and layout, expensive, slow, and frustrating, then fed what he learned into GPT-5 or 6 Pro and built workflows for book two. When he was done he cut about 14%, mostly by hand with AI helping decide what to drop, especially the over-explanation models do badly.

He has heard of Amazon romance writers shipping a couple of hundred AI books a year, a scale game in which a few titles make a few dollars and the pile is profitable. Custom artwork as good as what he spent months on will be available on demand. If a new Bergman or Kubrick film could be generated just for you, and shared culture is already fragmenting, is that still art? He does not know. He thinks we are in a sweet spot where a work you are proud of still takes a great deal of human judgment, and he does not know how long that lasts. The satisfying loop, for him, is to hand the whole novel to a model, get back an analysis and a list of things to consider, and decide point by point, like an assistant, rather than grinding the pass manually.

Borgan is also a musician, and he had held out a test for well over a year, maybe two: a high-definition capture of five notes on viola, given to every reasoning model. Clef, notes, values, time signature. Not a single one got it right until Astra, after OpenAI had presumably put actual scores into the training mix. Then he gave it a painful twentieth-century page, double sharps, accidentals, ties everywhere, and Codex did an almost perfect job turning the PDF into something machine-readable. Zero to 99 overnight. That is how you convert the mountain of unscanned classical scores into training data, and he would love Suno to ingest large-form work, operas and half-hour movements, not only pop songs. Private art for one listener still has value. He would still want to make it in tandem, with his taste in the loop, for as long as that role lasts.

Daniel McKinnon founded Gammo Labs to put AI agents on genomes and reopen cases that had been called unresolved. His son Owen died of a rare lung disease. A human specialist later found the genetic deletion the original sequencing analysis had missed. Last summer, during a closely watched pregnancy, a sixteen-week anatomy scan showed something his maternal-fetal medicine doctor would not normally have flagged. A fetal whole genome came back negative. It was the third negative whole genome he had seen. Between Owen and their now-healthy thirteen-month-old, Warren, they lost a second pregnancy very late for genetic reasons. Five years of heartbreak, all of it genetic mystery. He was heartbroken and mad, and he had tools. He called the labs, invoked HIPAA, demanded the raw data, and vibe-coded an interpretation pipeline on what was then autocomplete-era Codex, which still took real work. He was shocked that it outperformed on other cases, and most clearly that it diagnosed Owen when one of the best prenatal sequencing labs had not. He knew then he had to contribute. He did not yet know it would be a company.

The deletion was 91 kilobases and removed an enhancer, a region that helps regulate a gene. The enhancer sat a megabase, a million bases, upstream. What Rady did, and what many labs do, is filter structural variants, which are messy in 150-base-pair next-gen fragments, to within one kilobase up or downstream of a gene. That is a reasonable trade-off if humans are looking. McKinnon's question, before Codex and Claude Code, was whether you could put o3, the first real agentic breakthrough after o1, in a loop and just have it keep looking. The first pass asked whether anything was wrong with the coding elements. It missed, looped, and kept working. A clinical lab, profit or not, has to spend a finite time on each case and then call it nondiagnostic. Most genomes, even from infants suspected of genetic disease, come back that way. Export the looking onto a machine that can work nonstop and in parallel, and you see more kids. He does not claim to have invented genomic software. He claims to have noticed early that the task is long-horizon, agentic, verifiable, and unsaturated, a math-olympiad-shaped hill you can keep climbing until it is basically solved.

Gammo only sees hard cases, reanalyses that did not previously have an answer. The majority land in VUS, variant of uncertain significance, scored on an ACMG rubric. You need six points or more to reach likely pathogenic, which unlocks treatment and insurance. Stuck in the middle is a bad place. One way out is waiting until more patients with the same phenotype and variant appear, the epilepsy-for-ten-years story that sometimes ends in an off-label drug. It is not scalable. The other fork is convincing a university lab to run a functional study in a physiologically relevant cell line. For alveolar capillary dysplasia that line is IMR-90, fetal lung, because the gene is only expressed from week sixteen to week twenty of development; a cancer line will not do. You edit, insert a plasmid, and ask whether the gene is broken. Functional evidence can be worth two to four points. You only needed six.

The team bought what remained of Arpeggio, hired four people, and pivoted to making RL data from real biology. Three weeks in, they were running the first experiments the day of the taping. McKinnon wants the loop closed: VUS, then a study, then a region of the genome marked as important, then better predictions. Clinical genetics labs want the service too; it is not commercially available. Sequencing, he says, is no longer the bottleneck. High-throughput labs announce genomes for less than $100. Interpretation is. Twenty-five years after the White House genome celebration, a graveyard of genomics companies sits on that gap. o3 last summer was the first model that made the interpretation layer, and everything around it, explaining to a provider, ranking variants, ranking patients for ASO eligibility, feel possible. On RareBench, the best traditional machine-learning ranker, LIRICAL, scores something like 10%. Opus 5.5 in Claude Code, vanilla, scores something like 50%. The product he wants is cheap and usable by people who are not sitting in a top-tier hospital surrounded by counselors.

There is already product-market fit for sequencing sick babies in top NICUs, and lower-tier NICUs would sequence if they had the resources. The scoped question, baby has pulmonary hypertension, figure out why, is the one Gammo is working. One fork is online: a case comes in, traditional labs failed, the patient consents, data is reanalyzed, and sometimes, not always, the result changes a pregnancy or a newborn decision. The other is bulk. Rare-disease centers hold thousands of genomes and a handful of counselors. A machine can turn around 80, 100, or 200 in a week, six months to a year of work they cannot prioritize. The company is four months old and pre-revenue. They are not optimizing a business model. They are trying to diagnose more sick kids.

What does Gammo actually do? McKinnon, who started Meta's first LLM project, OPT-175B, and later worked at Google, will not bet against the frontier labs at the model layer unless you are famous and deep-pocketed. Vertical AI, he thinks, is harness, tools, and evals. They ensemble. RareBench's confusion matrix clusters: Claude is good at this, Astra at that, even Gemini at a third, and smart routing is an edge. Evals have to measure the whole system, not the model, with trace analysis with and without the harness. Not that many people look closely at the data. Grok 4.6, then state of the art on the benchmark, missed a case because it renamed the gene, sliding from something like NRF2 to a totally different name. The harness is there to stop dumb things. With millions of rollouts the labs might converge anyway, especially now that models write their own tools. Ten dollars a case, even $100, is fine. A million rollouts that make the case $1,000 or $10,000 is not. The job is to mix the smartest intelligences, give them the best tools, get answers for patients, and measure whether it worked.

Prakash asks how they evaluate the tools as model families change. McKinnon does not have a clean answer; they codesign. Astra, on a literature-search tool, said it had read the tagged papers. The answer was in a paper. Looking at the trace, they did not think it had actually read them. The first time they ran the Astra eval, the model stopped because some tools made it want to think too much, the same failure mode as a research agent that wanders off into Emirates flights from Dubai to Calcutta. They caught it early, changed how Astra called the tools, and it worked again. Paper cuts, stacked into rails. He wants the labs to hill-climb his task. He is both bullish and bearish on vertical AI: owning the customer relationship has value, OpenEvidence-style, and the layer is getting thinner. Vanilla o3 scored 0% on RareBench when he re-benchmarked it for fun. Frontier models now put up something like 20 or 30 points. The scaffolding he had to build is shrinking. The goal is still an AI-native clinical diagnostics company, and he wants the best intelligence doing the work. Cheap biology at scale, he adds, is arms, plates, cartridges, primer design, and analysis, a confluence, not one invention.

Talking to the frontier labs about the bio use case has been, for him, really positive. Nobody asks why he is doing this. Everyone wants to help, which is not the same as pivoting a trillion-dollar company, a distinction he understands from the inside. He would be surprised if the top labs spent no time on the problem. It is meaningful, and selfishly they need the story. Dario tweets about curing cancer in five years; Gammo is diagnosing kids today, four months in, at a small company, with multiple examples of children who had been undiagnosed. That is a concrete answer to people unhappy about data-center build-outs and jobs. Mid-career engineers who have spent a life on ads ranking hear that they can help kids and want in. After the interview, Labenz puts the trade-off on the table. The 50% is a RareBench score, not a cure rate. He still thinks pacing the frontier is probably worth it, and he does not think the labs will fail to grow if they do. He also will not take lightly that half the questions remain unanswered, and that those questions belong to individual families. It is a costly compromise even if, on balance, he would still make it.