Summary

The AI-bubble burst window is October 2026 to March 2027, if the Iran war continues and produces a correction. Murdock’s mechanism centers on credit-market disruption: hyperscalers have more debt than ever, private-debt spreads are “too narrow between real risk and not so much risk,” and complacency is the biggest warning sign. Japan’s Treasury holdings are a sleeper fuse: a $100B sale could be absorbed, but $300B to support the yen would create “an immediate global problem.”
“At least half” of the neoclouds go away within 36 months—faster in a dislocation—while hyperscalers are best prepared to survive. They can acquire cheaper assets after weaker players are wiped out, while AI-compute demand remains intact. What separates neoclouds is management quality, which outsiders cannot see; Murdock favors Fireworks over Baseten as a “10-times-better business” because of capital efficiency and a greater willingness to make profits. He believes Baseten’s Cursor contracts generated revenue and scale but little profit.
Open source and ASIC chips are a “tsunami of their own,” and tokens are not fungible. Against Gavin Baker’s “a token is a token,” Murdock argues that customization changes a token’s value. He expects frontier models to capture most dollars early while open source catches up and fills unmet demand, with short-term disruptions possible. Continuous-learning models could eventually replace today’s models, which also limits the long-term significance of backdoors in current Chinese or other open-source models.
Security is the most underestimated layer: “if you don’t get the sandbox right, forget everything else.” Harry cited Anthropic saying its models had hacked three companies; Murdock’s response was that containers are not safe and agents are probabilistic. An agent might open 100 sandboxes with 100 libraries to determine the best result. E2B and Docker are, in his view, probably the two strongest companies in this area.
OpenRouter’s 5% inference markup “is not gonna last.” Murdock points to exchanges such as Akinaki’s DODEx on mainnet and Venice, which could enable direct inference purchases and disrupt the model within three to five months. But if a hypothetical $10B Stripe bid arrives, his answer is “fuck yeah”—take the money. His Flipboard lesson is that refusing an approximately $1B opportunity can leave “a lot of arrows in my back.”
For margins, land grabs are acceptable as a strategy but not as a culture. Harry cited roughly 35% margins at Fireworks and around 20% at many AI application companies. Murdock favors companies that can eventually monetize innovation rather than simply buy customers. He also favors niche specialists over broad legal platforms such as Harvey and Legora, because a security failure could damage both.
Mag7 verdict: hold Meta, Google, and Microsoft long-term, but short Meta if forced. Meta and Google’s huge user bases—and Microsoft’s enterprise and consumer businesses—act as buffers. Microsoft’s Exchange business is a “money machine that cannot change.” Meta may become boring like AT&T, but Murdock still sees it as a stable, dividend-like holding. NVIDIA is over $10T in five years; its current plateau partly reflects circular transactions obscuring real growth.
His five-year contrarian call is blockchain for agent payments. Bitcoin’s perceived greed and hacking risk have dragged down the sector, but he sees long-term potential in Solana and Ethereum and real utility emerging through tokenized assets, payment rails, and inference systems such as Aki-Naki and Gonka.

Deep dive

1. The burst window: October 2026 to March 2027, on credit complacency

The forecast discussed is that, if the Iran war continues to fester and produces a correction, the AI bubble could burst “between October 2026 and March 2027.” Murdock’s historical mechanism is that financial disruptions interrupt commerce and slow innovation: the 2001 collapse delayed the next wave until the LAMP stack and Google, while the 2008 crisis slowed cloud adoption.
This cycle is unusually debt-heavy. Hyperscalers have taken on more debt than ever, creating vulnerability if the credit markets or broader capital markets are disrupted. Murdock repeatedly frames complacency—not one decisive data point—as the central warning sign.
Private-debt spreads are, in his view, “too narrow between real risk and not so much risk.” He compares the situation with 2008, when credit-rating agencies and bank risk departments were “asleep at the wheel.” He also cited a leveraged fund or person transcribed in the reference as “Leopold [?]” as a recent warning about leverage that was not properly accounted for; he later said the entity was 3.5× levered and had to sell assets.
Japan is another potential fuse. Murdock says the U.S. has bailed out the yen twice because Japan holds $1T in Treasuries. A $100B sale might be absorbed, but a $300B sale to buy dollars and support the yen would create “a real problem on our hands immediately.”
His backcountry-skiing analogy is that experienced skiers recognize avalanche conditions: the issue is not that collapse is certain, but that several forces can make the system easy to tip. A worsening war, renewed inflation, or a broader market break could all become disruptors.

2. Hyperscalers survive the dislocation; half the neoclouds do not

Harry’s pushback was that today’s assets are better than the weak dot-com companies of the past: Meta has a strong core business and, in Harry’s framing, does not truly need to borrow. Murdock replied that Meta’s free cash flow is “the lowest it’s ever been in the history of the company.”
He also used dot-com fiber as a counterexample: the fiber itself remained valuable, but the companies that laid it went bankrupt. In a dislocation, heavily debt-dependent companies can see the value of their assets decline sharply in a short period, producing margin calls even when the assets retain long-term value.
Nevertheless, “no one is better prepared to survive it than hyperscalers.” Their ongoing businesses are consistent, AI-compute demand does not disappear, and a dislocation could let them acquire assets more cheaply while weaker competitors are wiped out. The near-term problem is funding, not demand.
Murdock expects at least half of the neoclouds to go away within 36 months, with many disappearing immediately if an economic disruption arrives. The differentiator is the people running and organizing each company—information outsiders cannot see “under the covers.”
His concrete proxy is Fireworks over Baseten. He said Fireworks was making much more money, called it a “10-times-better business” in his opinion, and attributed the difference to capital efficiency and a willingness to make profits. He believes Baseten’s contracts with Cursor from the prior year delivered revenue and scale but probably not much profit; putting up substantial capital without generating earnings creates risk.

3. Open source, ASICs, and specialized intelligence

Murdock maintains that open-source models and ASIC chips are “a tsunami of their own.” He says companies cannot currently customize the large Anthropic and OpenAI frontier models, creating an opening for tuned open-source models.
His cost comparison is a frontier model with a double-digit cost per token versus an open-source model at roughly 10 or 11 cents per token. He stresses that tokens are not necessarily equivalent, but considers the difference large enough to drive massive adoption. He estimates current global AI-demand fulfillment at the low single digits.
He expects frontier models to receive most early revenue because wealthy companies and people can pay for them, while open source plays catch-up on revenue and fills a large unmet need. He also allows for short-term disruptions lasting three months to a year in which the economics appear to level out. If frontier companies achieve continuous and ultimately lifelong learning, however, he believes demand for them will continue to grow rather than being permanently cannibalized.
Against Gavin Baker’s “a token is a token,” Murdock argues that customization changes a token’s value. Different models may be verbose or brief, and a customized model can perform a specialized task more efficiently. He says a company with only $1M to spend may get more value from customization than from spending the same amount on a frontier model.
He expects specialized models to resemble specialized human intelligence: a gem cutter, for example, has a specific kind of expertise. Open source may be especially suitable for coding, customer service, and onboarding, where narrow specialization can pay off quickly and cheaply. He says open source has not yet surpassed frontier models in innovation or complex tasks, but can be better at specialization—at least for now.
ASICs are attractive for customized models because specialized tasks do not require an expensive GPU. Murdock sees owning chips as useful for large companies in the short term to optimize for their models, but unnecessary in the long term. His broader investment focus is the complexity between model, agent, and human, including customization, security, and the loops between them. He also recalled a Santa Fe Institute takeaway that nobody knows how to measure AGI even if it appears.
On data, Murdock says it is not static: it provides context and memory, and its value depends on how each enterprise uses it. Shake Shack, Burger King, and McDonald’s may all have data about hamburgers, but their data has different applications. Systems therefore need to evolve as both the data and its use change. He expects AI to move from task completion toward deeper creativity and a diversity of intelligence capable of addressing difficult problems.

4. The security choke point is the sandbox

Harry cited Anthropic saying its models had hacked three companies and characterized that kind of disclosure as almost a brag. Murdock’s response was that security complacency is widespread: developers may put a model and its tools in containers and assume they are safe.
Murdock says Docker itself has warned that containers are not safe on their own, which is why he points to Docker Sandboxes and E2B’s cloud sandboxes. His conclusion is categorical: “if you don’t get the sandbox right, forget everything else.”
Agents are probabilistic rather than deterministic. One might open 100 sandboxes with 100 different libraries, test them, and select the best application. Murdock says only a handful of companies understand how models interact with tools and how to optimize for that behavior; E2B and Docker are probably the two strongest examples.
On Alex Karp’s warning that major enterprises may avoid frontier providers, Murdock says Karp’s CIA and intelligence-agency customer base is particularly concerned about security. At the same time, he argues that enterprises have already given substantial data to third parties: Apple has extensive personal information, Amazon has significant data, and Satya Nadella has described Microsoft’s knowledge of organizational communications.
His prescription is discernment rather than panic. Enterprises should decide what remains behind the firewall and should not continuously upload sensitive data to Anthropic or OpenAI without considering the consequences. That concern also creates an opening for open-source and privately deployed systems.

5. Land grabs, margins, and hype-cycle pricing

Murdock compares early-cycle pricing to the Oklahoma land rush: companies may accept low or zero margins to plant their flag, win customers, and secure the relationship before building margins later. But he calls this a strategy, not a culture. He would not invest in a company that is structurally comfortable with low margins.
Harry cited Fireworks at roughly 35% margins and many AI application companies at around 20%. Murdock said a sandbox company should ideally let customers bring their own compute and charge for its expertise in operating, networking, tracing, and securing the sandboxes rather than simply reselling compute.
He says innovation should eventually drive margin. He attributes his having probably missed investing in Amazon to not accepting Bezos’s “your margin is my opportunity” approach, rather than treating that strategy as generally appropriate.
On Harvey versus Legora, Murdock recommends funding the smaller startup and watching the well-funded competitors battle. He expects a security leak to occur and potentially damage the first company hit, but warns that both may be vulnerable because they are focused on each other. He prefers highly specialized businesses such as GetDynasty in trusts, consistent with Peter Thiel’s advice to dominate a niche before expanding.
Harry described founders treating $100M rounds as friends-and-family financing and $25M checks as small and collaborative. Murdock views this as evidence that the market remains in a hype cycle. He thinks the Anthropic and OpenAI deals he described at $100B and $150B may have been cheap, but insists that investors must distinguish genuine hyperscale opportunities from also-rans.
When Harry asked whether “triple, triple, double, double” still applies, Murdock called him glib in this case and rejected general rules. He said frontier companies have done something extraordinary—“on the level of inventing fire”—and that infrastructure directly supporting them may deserve those economics. He does not extend that conclusion to application companies or neoclouds.

6. OpenRouter’s 5% markup and knowing when to hit the bid

Murdock says OpenRouter has massive transaction volume because developers are willing to pay for convenience, but its 5% inference markup “is not gonna last.” He points to Akinaki’s DODEx on mainnet, Venice, and other exchanges being built to let customers obtain inference more directly.
He expects a major disruption to that model within “the next three, four, five months.” If OpenRouter receives a hypothetical $10B acquisition bid from Stripe, however, his advice is “fuck yeah”—take the money. OpenRouter arrived early, solved a real developer problem, and earned the markup before alternatives were obvious.
Cursor is another example of a potentially non-repeatable liquidity event. Murdock says the team pivoted out of the IDE space, convinced Elon that it could build models before proving it, and then received a $60B outcome through xAI. If OpenRouter receives a $10B bid, he would similarly “hit the bid.”
Flipboard is his counterexample. It had two bidders—Twitter and TikTok through ByteDance’s founder—interested at a figure within roughly 20% of $1B. The board followed advice from “The Coach,” who was then passing away, not to sell. Murdock says the missed opportunity left him with “a lot of arrows in my back.”
He does not see Airtable at $4.85B as proof that an entire generation of companies will be sold at a discount. There are relatively few buyers such as Bending Spoons, and the real question for a $400M–$500M-revenue company is whether that revenue will persist. If a SaaS company has not developed a compelling AI strategy, he is not optimistic about its prospects in two years.
He says Cursor could have gone public the prior summer, since there is always a banker willing to attempt an IPO, but management correctly judged that it was not ready. The availability of a market is not the same as an IPO being the right decision.

7. Co-work, levered private equity, and venture discernment

Murdock calls the shift toward autonomous agents the “co-work era.” Copilot-style additions may not protect SaaS companies that lack a system of record or a serious AI strategy, though he says there is still time to pivot.
Private-equity leverage is the danger in a dislocation. Harry cited assets at roughly 4–6× leverage; Murdock noted that the entity transcribed as “Leopold [?]” was only 3.5× levered and still had to sell assets. If EBITDA falls and churn rises quickly, debt can create an effective margin call.
He recalled TPG’s difficult 2001 fund, which survived, versus Forstmann Little, whose telecom exposure helped end the firm. He also recalled Tom Lee—“I think it was Tom Lee”—calling for a 10% S&P drawdown in the fall; if the decline is worse, Murdock is unsure how highly levered assets will cope. He feels best about Insight’s long-term position because its private-equity portfolio is very small, not because he thinks heavily PE-focused firms are insulated.
His general venture advice is to look for people who are different from everyone else, whose businesses would have real impact, and who feel they have to build the company rather than merely want to build it. He gives Fireworks and AtoB as examples, and says founder-driven companies such as Elon’s, Jensen’s, and Zuckerberg’s fit that definition.
On Harry’s premise that Sam Altman could give the administration 5% of a frontier-model company, Murdock says government ownership has not historically been necessary in the United States and that, given the company’s existing scale, the rationale would be political rather than a Manhattan-style strategic program.
On chip export controls, he does not offer a simple yes-or-no endorsement. He calls instead for a coherent technology strategy that is public, debated, and assigned to responsible decision-makers, rather than ad hoc regulation.

8. Continuous learning and the Mag7 quickfire

Murdock says current open-source models will not be used in ten years, and that continuous-learning systems may arrive in two or three years—or may take ten. These systems would require new architectures and training, ultimately replacing today’s models, including current frontier models. He does not think continuous learning can simply be bolted onto an existing frontier model.
That is also why he treats current Chinese-model backdoor concerns as time-limited: the current models may disappear before long. His confidence is hedged. Sample-efficient models are beginning to learn from small amounts of data, but he says a larger breakthrough is still needed. A global financial event combined with failure to advance continuous learning could send the sector into the “valley of disillusionment.”
He declined to judge SSI at $30B because he does not invest in model companies unless he knows the founders or someone he trusts knows them.
Quickfire: he said “it appears like Anthropic” will go out before OpenAI. He expects NVIDIA to exceed $10T in five years. Its current plateau reflects the fact that markets do not rise like a rocket ship forever, uncertainty around demand, and circular transactions that may be obscuring organic growth.
Asked to “shag, marry, kill” Meta, Google, and Microsoft, he chose long-term ownership for all three. Meta and Google each have roughly two billion users, while Microsoft has a strong enterprise and consumer business; those installed bases provide buffers and time to recover from AI mistakes. He says all three have failed so far on coding agents, but may not fail permanently.
Forced to short one, he chose Meta because it may become boring, like AT&T. He still considers its WhatsApp, Instagram, and advertising scale stable enough that it could become a dividend-like safe haven. He remains uncertain about Apple: it is a major Claude Code customer, but if it is merely consuming others’ innovation rather than observing and building something of its own, it will suffer.
Among venture firms, he names Khosla, Menlo—early enough into Anthropic—and Benchmark, which missed Anthropic and OpenAI but performed strongly with Factory and other companies.
His five-year “crazy today, obvious later” prediction is blockchain for agent payments. He says Bitcoin’s greed element and warnings from IBM’s CEO, Tom Lee, and Google about possible hacking in roughly two to four years have dragged blockchain into the trough of disillusionment. Solana and Ethereum appear to have long-term potential, while newer systems involving tokenized stocks, payment rails, and AI-inference exchanges—such as Aki-Naki and Gonka—could demonstrate genuine utility.

AI in the AM — Weekly Highlights: Relaunch Week (Aug 17–20, 2026)Simulating Humanity: from Generative Agents to 8 Billion Digital Twins — Joon Sung Park, Simile AI