OpenAI is Making a MASSIVE Mistake

Summary

• Travis Good’s headline call: “open source is going to win and you still have time to be the hero” — the closed labs’ maximum-extraction posture can destroy the one asset they actually have, distribution. His warning to Sam Altman and Dario: “your distribution doesn’t matter because your customers hate you,” triggered by Anthropic revealing — buried “somewhere in their 300-page release on Fable” — that it sabotaged ML-research queries, then “walking it back” to silently falling to a less capable model, “which is also sabotage.”
• Ambient’s pitch is “Uber for inference” running on Costco economics: it owns no GPUs, just the network that routes every request and guarantees quality. Roughly 25% of the world’s enterprise GPU capacity sits with small operators earning ~10% rental margins, some potentially against 17% loans, while serving inference earns 40-50%. Ambient’s verified network and demand aggregation let them pursue that uplift without their own business-development arm.
• Verification is the moat, and Good says it’s now “100% logits” and “basically costless,” toggled per request. His anecdote about an aggregator’s weekly intelligence tests: providers slow to ~50% of normal speed on test day, apparently switching to a full-fat model for testing and a compressed version otherwise. “If you could get verified and unverified inference for the same price, why would you ever choose unverified inference?”
• Miner economics stack a transaction payment with inflation rewards that work “almost like stock-based compensation,” so bullish operators can underbid 10-15% per transaction and still come out ahead. Ambient’s routing evolved from a pure auction doing 10M tokens a day to a cache-aware system at 100M and then a “predictive HFT-style algorithm” doing 10-15B tokens per day per operator. It is already typically #1 on OpenRouter serving Kimi K2.7 — ahead of Moonshot on that measure.
• The “China goes closed-source” bear case gets a judo answer: training costs have collapsed (GLM-5.2 for ~$25M, perhaps $5M next year), so any lab that closes could be challenged by new open-weights entrants. Three more years of this could make pre-training so commoditized that value shifts toward delivery and supply aggregation — exactly Ambient’s position. On the distillation debate, Good says China is “largely innovative” and distillation is “almost incidental” because “we’re all distilling each other.”
• In the hard-takeoff scenario, Good says only the US government would get access; Tommy argues that no company could then create economic productivity, and that a limited public version would simply fall back to Opus 4.8. Tommy warns that distributing only the model’s “golden apples” would put the government in the business of picking winners and create a centrally planned economy. Good agrees he does not want that outcome; Tommy concludes that “the diffused world is a lot more functional.”
• Tommy calls unsolved prompt injection both a “trillion-dollar business opportunity” and a potential handicap for closed-source AI, creating a “regulatory prisoner’s dilemma.” His concern is that safety-conscious America could withhold models and still get “hacked six ways to Sunday even though we have superior model capabilities,” while Anthropic, the safety poster child, still cannot solve prompt injection. Good engages the concern and notes that restrictions could let competing countries release better public models.
• The capex unwind scenario: Good imagines OpenAI raising at $500B instead of $1T, failing to fulfill Oracle commitments, data-center buildouts stalling, and “what happens then is probably a 2-year AI winter” — with 401(k) hits and anti-AI backlash. Tommy adds that per-million-token prices are down 20-25% in 30 days as businesses swap frontier APIs for models offering 90-95% of the intelligence at a fraction of the cost — a potential opening for asset-light aggregators, but a threat to the broader funding cycle.

Deep dive

1. Anthropic’s Fable sabotage is the case study for credibly neutral infrastructure

• Travis Good’s read on the Alex Karp video: the pain everyone just felt is “what happens when your trusted provider turns against you.” Anthropic disclosed — without emphasis, “somewhere in their 300-page release on Fable” — that it was sabotaging ML research queries, then walked it back to falling to a less capable model, “which is also sabotage.” Tommy’s reaction: “feel rugged on that one.”
• The Karp point Good endorses: “if you’re competing with Anthropic you are in their crosshairs… you don’t want to be in the crosshairs of an opinionated infrastructure provider.” That conflict of interest is precisely the opening for crypto and credibly neutral infrastructure.

2. Uber for inference, Costco underneath: matching stranded GPU supply to guaranteed-quality demand

• Good’s business framing: “Uber for inference — we don’t own the cars, we own the network that matches every request to the best available supply and guarantees the quality,” focused on two canonical models, the most popular high-intelligence large model and its small-model counterpart. One layer down it’s Costco: “suppliers fund Costco’s inventory, but Costco turns it fast,” keeps prices low, and owns the customer.
• The supply gap: about 25% of the world’s enterprise GPU capacity sits with operators earning poor margins — 50-100-GPU shops stuck renting at ~10% margins when inference service earns 40-50%. Ambient is the trusted brand that guarantees delivery quality so they can capture that uplift.
• On availability, his exhibit is the Dubai AWS data-center bombing that took Anthropic down across the southern US: providers are “operating at the edge of their capacity envelope,” while Ambient’s continuous auction plus surge pricing means no arbitrary rate limits — unlike Together, which he says is largely booked up.

3. The contrarian design: focus extreme economic attention on a few models

• Good’s self-diagnosis of the industry: “crypto has lacked a certain kind of ambition” — the ambition to go head-to-head with Web2 business models. Traditional providers are heavyweight and inherently biased by their leases: “I need to feed my book first,” so they can’t credibly arbitrage supply, while Ambient is asset-light, nimble, and antifragile.
• The Bitcoin lesson behind serving few models: concentrate economic attention and “10,000 people working frantically to use whatever hacks they possibly can” solve delivery for you — the same CPU→GPU→ASIC progression, since Ambient is also proof of work.
• His shopping-mall analogy: be the anchor store — the Dillard’s or Nordstrom — so efficient at the canonical models that “you wouldn’t want to do a margin game with us,” which then draws suppliers and customers into higher-margin boutique activity around it.

4. Verified inference is now 100% logits and “basically costless” — and one test anecdote suggests providers may be gaming quality checks

• The problem is “intelligence compression”: “Claude was dumber this week, OpenAI was dumber this week” — and it happens with open weights too, since serving patterns are highly variable. Without verification, quality is inconsistent and the business isn’t scalable.
• The algorithm’s evolution, as Good tells it: a year ago it was three levels — logits, internal model telemetry, complex mathematical relationships. Today it’s “100% logits,” a fingerprint of the model’s thinking — choices made, the spread not taken, associated probabilities — secure, per-request toggleable, with no visible speed difference, across vLLM, SGLang, and a llama.cpp beta.
• His anecdote: an aggregator Ambient works with runs weekly intelligence tests, and serving speeds drop to ~50% on test days — everyone apparently serves the “full-fat version” for the exam and a compressed version the rest of the week. His standing question: “if you could get verified and unverified inference for the same price… why would you ever choose unverified inference?”

5. Miner economics: transaction rewards plus inflation “stock,” and routing that went from 10M to 15B tokens a day per operator

• The cheeky answer to “why would anyone optimize for Ambient”: “we’re doing better than Moonshot at serving Kimi K2.7 on OpenRouter today” — typically top three, usually number one. The real answer: stranded operators with 17% loans making 10% margins have no business-development arm fighting for demand; 10% to 40% margins “is transformational.”
• The token design doubles the incentive: every job pays a transaction reward plus inflation-based rewards — “almost like stock-based compensation” — tied to contribution and epoch share. Believers can come in “maybe 10 or 15% under” on price and still realize large overall gains. “I’m not getting shares in Together… they’re just taking their cut.”
• Routing went through three iterations: a pure auction doing 10M tokens/day on OpenRouter (“we were sad about that”), then cache-aware routing at 100M, and finally “an incredibly complex high-speed predictive HFT-style algorithm” — uncheatable because it relies only on external metrics — that routes exactly what each miner can take without tipping over, now at 10-15B tokens/day per operator. “I still think there’s gas in this tank.”

6. Beyond inference: a “programmable machine-learning economy”

• For the brilliant researcher without capital, the Ambient Foundation is exploring pooled support — crowdfunding nodes with Ambient tokens and renting supply through its book of providers, so an optimization can reach the network without its author buying GPUs.
• The bigger reveal: “Ambient is a programmable machine-learning economy.” Inflation is a budget spendable programmatically — flip a portion toward miners doing training jobs for a period, share profit back to the network. The mall’s boutiques get hosted while the GPUs are “getting fed by delivering commodity models at high speed for good margins” day to day.

7. What if China closes the source? The judo answer is $5M training runs

• Tommy’s bear case: Zhipu and peers trade at “14,500 times sales” (as quoted) while US neoclouds serve their open-source models; one day China could close the weights or demand a revenue cut. Good agrees that closing is possible but flips it: a year ago it was “a radical heresy” that GLM-5.2 could nip at the closed frontier for $25M; if next year it costs $5M, investor capital could fund teams releasing open weights that are “trashing GLM-5.6 in the ratings” — “didn’t we see a Chinese DoorDash competitor release some ridiculous owl model?”
• The conclusion that matters for the thesis: if the dynamic runs another three years, pre-training knowledge could become highly commoditized and “it really becomes all about model delivery and supply aggregation — which is where Ambient sits.” The risk, ironically, creates opportunity for aggregators.
• On innovation versus distillation, Good calls China “largely innovative” and distillation “almost incidental”: “everyone is a 10x engineer” now, and perfected open artifacts — like his own MIT-licensed Ambient Desktop at over a million lines, built mostly in off hours — are ideal training data. “We’re all distilling each other”; you don’t need Anthropic’s user data when the pull-request history shows how Claude revised everything.

8. Direct message to the closed labs: give something away before your customers finish hating you

• The core declaration: “open source is going to win and you still have time to be the hero.” America’s soft power came from sharing culture and research; his challenge — “would it kill you to release some research on a six-month time delay?” — buys goodwill, bends global architecture America’s way, and wins mindshare. Gemini is popular but “getting a little long in the tooth”; imagine another OpenAI open-weights release. Right now “the world feels like it’s being beaten up by America’s closed AI.”
• Tommy pushes for the dollars-and-cents case, not altruism. Good’s answer: OpenAI and Anthropic hold “the world’s ultimate Rolodex” — a distribution advantage backed by massive venture investment — and can afford to give things away while capitalizing on it. The max-extract risk: “your distribution doesn’t matter because your customers hate you.” Taking customers’ data and disintermediating their business relationships “creates a fear response.”
• On the Palantir “own your weights, own your data” worldview, Good says Ambient can offer a private alternative — trusted execution environments, onion-routed requests, and PII stripping — versus an Anthropic enterprise call where “they twist your arm… and maybe won’t do ZDR for the latest models because of safety.”

9. Steelmanning the hard-takeoff bet — and why it still loses

• Layer one, the data centers: Good says everyone is asset-heavy in the NVIDIA GPU-cluster paradigm, perhaps including Vera Rubin, and calls that “an incredibly risky bet because ASICs are a thing.” Tommy develops the ASIC case: 14-nanometer chips using SRAM would not need high-speed HBM, could be produced at massive scale, and might offer vastly more capacity for the same spend. Good’s response is that Ambient remains asset-light; if a shock forces GPU dumping, those GPUs could find a home on its network.
• Layer two, the models: Tommy argues that trillion-dollar valuations only make sense if self-improvement produces an insurmountable edge, not a 3-6-month lead. “Everyone is a machine-learning engineer now,” he says; moderately skilled humans overseen by average-intelligence LLMs “might be just good enough” to keep pace.
• Tommy’s EV pushback — a 1% shot at a hundred trillion justifies burning the boats — gets played to its conclusion. Good says OpenAI achieves hard takeoff and “nobody in the US except the government gets to access it” (hack the NSA, get banned). Tommy argues that without company access there is no economic productivity, that a limited public version would simply fall back to Opus 4.8, and that distributing only the model’s “golden apples” would put Washington in the business of picking winners: “we become the China we’ve wanted to avoid.” Good agrees he does not want a centrally planned economy, while Tommy concludes that “the diffused world is a lot more functional.”

10. Prompt injection: the trillion-dollar hole in the safety story

• Tommy’s evidence of misallocated resources: safety researchers gave us P(doom) and cyber-war scenarios, “and yet we have prompt injection.” Solving it is “a trillion-dollar business opportunity” because it could enable everyone to use full-fat Fable; without it, even partially AGI-capable models might be unreleasable, a friction he says Leopold Aschenbrenner’s curve ignores.
• Tommy calls the geopolitical version a “regulatory prisoner’s dilemma”: the safety-conscious side withholds models and could get “hacked six ways to Sunday even though we have superior model capabilities,” while other countries release theirs. His jab is that Anthropic, the safety poster child since its creation, still cannot solve prompt injection. Good responds that restrictions could let China and other competitors release better public models.
• On harnesses, Good says “Claude Code is pretty bad as a harness; Codex is pretty good. Hermes is way better than OpenClaw.” Tommy synthesizes that an open model plus company memory, data, and orchestration is a de facto fine-tuned experience without changing weights. He also argues that solving continual learning could make prompt injection “1,000 times worse,” since a poisoned input 10 million tokens earlier might later turn the model rogue; Good jokes that models would “continually learn how to prompt-inject.”

11. The two-year AI winter scenario — and Ambient’s launch plan into it

• Good agrees that this is the open-weights inference moment — Together AI, Fireworks, Modal, and Baseten each adding roughly $5B of value in five months — but the circular capex web involving OpenAI, Oracle, CoreWeave, and NVIDIA is fragile.
• Good’s chain: OpenAI raises at $500B instead of $1T and cannot fulfill its Oracle commitments; Oracle, which he says has “bet the company” on OpenAI, then cannot fund data centers. Tommy adds that Oracle could be marked down roughly 80%. The neoclouds drafting off frontier capacity hit a ceiling. “What happens then is probably a 2-year AI winter,” with 401(k) damage and an anti-AI populist backlash driving reflexive laws. It’s “the perfect storm for Ambient success,” but bad for America. Tommy’s near-term corroboration is that token prices are down 20-25% in 30 days as businesses take 90-95% of the intelligence at a fraction of the cost.
• The roadmap: a closed beta for small-model mining “in about the next week,” then an open PvP season anyone can join via the cross-platform Ambient Desktop app — including mining on a laptop. Pooled big-model nodes will run at 100% profit pass-through with no Foundation take, and the network launches this year.
• The one-year success picture: number one in delivery of 10 models on OpenRouter instead of one, mom-and-pop operators earning real margins, subscription and “X42” agentic-commerce adoption — becoming “the global intelligence utility that we’ve been promised by all the closed-source people but who are too conflicted to deliver that.” Good says Dario is never going to deliver intelligence as utility because his hands are tied by defense contracts.