Jerry Murdock on AI Financing, Open Models and Technology Cycles

Insight Partners co-founder Jerry Murdock joins Harry Stebbings to discuss AI financing, open models, security and the pace of technology cycles. The conversation includes market forecasts, company comparisons and geopolitical claims that are the speakers' views and are not independently verified in the interview.

Stebbings asks about Murdock's earlier warning that a prolonged conflict involving Iran could contribute to a market correction and an AI financing downturn. Murdock presents that as one possible scenario and time window, not a certainty. The conversation does not independently establish the course or effects of the conflict.

Murdock compares potential disruption with the dot-com decline and the 2008 financial crisis, when financing stress slowed parts of the technology economy. He sees debt as an important feature of today's infrastructure cycle. His historical comparison is a framework for risk, not proof that the same outcome will recur.

He worries that investors may underestimate credit risk and says private-debt spreads can fail to reflect differences in underlying borrowers. He points to complacency as a warning sign. The interview does not independently assess current spreads or demonstrate that a particular lender is mispricing risk.

He argues that both the cost and availability of credit matter because AI infrastructure requires substantial up-front investment. If lenders become more cautious, projects could be delayed even when demand for compute remains. He does not present an analysis of the financing terms of every hyperscaler or cloud provider.

Murdock recalls the failures of risk assessment around the mortgage crisis and argues that problems can be missed when people rely too heavily on past stability. He says several kinds of leverage could create vulnerabilities today, although he does not establish that the present credit system matches the conditions before 2008.

He raises Japan's holdings of US Treasury securities as a hypothetical channel through which currency stress could affect markets. His figures, characterization of past interventions and suggested market impact are his account and are not verified in the conversation. The scenario depends on decisions and market responses that remain uncertain.

Stebbings counters that large technology companies have valuable underlying businesses and substantial cash generation. Murdock agrees that hyperscalers are comparatively well placed to survive a disruption, but says even useful infrastructure assets can fall in value when debt markets tighten. He uses fiber networks after the dot-com cycle as an analogy.

Murdock compares infrastructure to fiber laid during the internet boom: the underlying asset can remain useful even if its original owner experiences financial distress. His point is that asset utility does not eliminate the risk attached to the way it is financed. This analogy is not evidence that current firms will fail.

In his scenario, well-capitalized hyperscalers might continue operating and acquire assets more cheaply after weaker competitors struggle. He expects demand for AI compute to persist but says short-term financing can still be disrupted. This is a conditional forecast, not evidence that a credit event or acquisition wave will occur.

Murdock predicts at least half of new cloud infrastructure providers may not remain viable or independent within three years, with faster losses if financing is disrupted. The percentage is speculative, and he does not establish the status or balance sheet of any specific provider.

Asked what separates survivors, he emphasizes management, organization, capital efficiency and the ability to earn a return. He compares Fireworks and Baseten favorably for the former while disclosing that he invested in Fireworks. His claims about their relative profitability and contracts are not independently confirmed.

He says he would favor a provider he believes can balance growth with profit and capital efficiency. His Fireworks investment is relevant to interpreting that preference. Without verified financial statements, listeners cannot independently assess his comparison of two private companies' earnings.

The conversation moves to specialized models. Murdock expects a variety of models adapted to particular tasks, alongside frontier providers. He argues that enterprises may want to customize open-weight models when their workflows or data differ from those of other customers. This is a forecast about a developing market.

He expects lower-cost open models and specialized chips to meet some demand for routine work. Stebbings asks whether open models could handle more token traffic while frontier companies still collect more revenue. Murdock says the pattern may shift over time and that short-lived changes in price and capability can disrupt any simple split.

Murdock believes frontier providers can continue to attract demand if they deliver new capabilities, while lower-cost models may serve a broader set of users and tasks. He estimates that only a small fraction of potential demand is being met, but provides no measurement for that estimate. Future market size and the split between providers are uncertain.

Stebbings asks whether the largest model providers will continue to receive most of the dollars even as open-model traffic grows. Murdock expects both categories to find customers but acknowledges that relative economics can shift. His analogy to expensive early electric vehicles is illustrative rather than a demonstrated pricing trajectory for AI.

He disagrees with the idea that one token is interchangeable with another. A customized model may respond with a different level of detail, accuracy or brevity; those differences affect the total cost and value of a task. A token price alone cannot tell a customer which system performs a job best.

As examples, he suggests onboarding and customer service could benefit from specialized models. The cost and effectiveness of such systems depend on data, evaluation and implementation. His view that customization will become widespread is a prediction, not a guarantee that every open model will outperform a frontier model.

Murdock says a model's style and verbosity can affect how much output it produces, while customization may change how well it fits a particular workflow. Stebbings interprets this as an efficiency advantage from adaptation. Both describe a possible effect; the outcome should be measured for the specific workload.

Stebbings asks whether cheaper alternatives will cannibalize frontier providers. Murdock responds that demand for different kinds of intelligence may expand as new uses appear, allowing several approaches to grow. He also expects technology cycles to include periods of expansion and contraction; those are his market expectations.

The discussion turns to enterprise concerns about depending on model providers. Murdock says organizations should consider what sensitive information they send to third parties and what they keep behind their own controls. His comments express a risk-management opinion and do not establish that a named provider has misused customer data.

He observes that companies have long used outside cloud, communication and software providers, so data governance is not a wholly new question. He expects some enterprises to use private or open models for particular workflows. Contract terms, deployment choices and technical safeguards differ and require case-by-case assessment.

Murdock's advice is to revisit decisions about sensitive information as products and company requirements change. A firm may choose different arrangements for ordinary work and more confidential tasks. The conversation does not analyze any particular model provider's current contractual terms or data practices.

On security, Murdock worries that teams may give agents too much freedom without adequate isolation. He says sandboxes can reduce exposure when executing tools or code. The strength of any isolation depends on its design, configuration and threat model; the interview does not prove containers or any named product are universally safe or unsafe.

He discloses his investment in Fireworks and praises its team, then points to sandbox providers as examples of an area he considers important. He expects agents to run many experiments or use many tools, making visibility and isolation valuable. His company rankings and forecasts are investment opinions, not a comparative security audit.

He argues that agents can run multiple tool calls and experiments, making it important to understand what they can access and how actions are isolated. He points to his preferred companies as examples, but does not test them against one another. Sandboxing is one possible layer within a broader security design.

Stebbings asks whether AI applications will have permanently lower margins than traditional software. Murdock says some companies may accept low or zero margins to acquire customers early, hoping to improve economics later. He treats this as a strategy rather than proof that every company can eventually raise its margins.

In his land-grab analogy, a provider with sufficient capital may accept weak early margins to obtain customers. But turning those relationships into future profits depends on continued value, pricing and competition. Murdock says he would still want to understand a credible economic path before investing.

He says he would be reluctant to invest in a company whose culture never considers how it will earn a margin. He acknowledges that he missed Amazon and that its approach to low margins and scale worked in ways he had not embraced. His preference is to identify innovation that can eventually support sustainable economics.

For sandboxes, he suggests a provider could differentiate through orchestration, networking and visibility rather than only reselling compute. Whether customers would pay for those capabilities depends on value and competition. He is describing a potential business model, not reporting the economics of a particular company.

The speakers discuss specialized chips for customized models. Murdock thinks dedicated architectures may be useful for some workloads and forecasts more activity in chip design. Suitability depends on model requirements, software support and total system cost; the discussion does not establish that one chip type is always superior.

Asked whether model companies should own chip design, he sees a short-term reason for large providers to optimize hardware for their models but doubts every company needs to own the layer long term. He finds the area between models, agents, human oversight, customization and security more interesting as an investor.

Murdock recalls a scientific discussion about measuring advanced intelligence and says the relationship among models, agents and people is a particularly interesting area. He points to customization, feedback and security as potential investment opportunities. His account of that discussion is not a consensus forecast for the industry.

Stebbings discloses an investment in Legora while asking about well-funded legal-AI competitors. Murdock says he might seek a smaller, more specialized company rather than bet only on the largest rivals. He speculates about security risks in a fast-moving category; he provides no evidence that a particular company has suffered or will suffer a breach.

He connects that preference to starting with a narrowly defined customer problem and expanding after building strength in it. A niche can provide focus, but it does not by itself guarantee a defensible market or an eventual exit. His choice is an investment perspective, not a conclusion about any specific competitor.

He also says a smaller entrant might focus on one legal specialty rather than attempting to serve all lawyers. Stebbings has disclosed an investment in one of the companies mentioned, and Murdock's preference is an investor's opinion. Speed and focus do not eliminate the need to protect sensitive customer information.

Stebbings describes very large early rounds and asks whether investors have lost price sensitivity. Murdock sees evidence of high expectations but says investors should distinguish exceptional opportunities from companies raising large sums because the market permits it. He thinks some past large rounds may prove inexpensive; that is his opinion, not a verified future return.

Murdock asks whether founders can explain why they need the amount they are raising, rather than selecting a number because large rounds are fashionable. Some capital-intensive projects may genuinely need substantial funding. He sees judgment about the business and the people running it as more useful than one blanket valuation rule.

They debate whether the traditional trajectory of several years of tripling and doubling revenue still matters in venture capital. Murdock rejects a universal answer. He views the current growth of frontier-model companies as unusual and believes infrastructure companies may grow differently from applications or new cloud providers.

He praises frontier-model progress as historically significant but does not assume every adjacent business merits identical economics. He remains skeptical of many new cloud providers and repeats his prediction that a portion will disappear or consolidate. The outcome for individual firms is unknown.

Stebbings asks whether model routing can support an independent business. Murdock questions the durability of a markup for routing and expects competition from exchanges and other distribution models. His characterization of OpenRouter's fees and potential alternatives is not independently verified, and future pricing may differ.

Murdock believes a convenient early routing product can attract developers, but expects new competitors to challenge its economics. He discusses an exchange model for purchasing inference as one possible alternative. The value of any service to customers depends on functionality, reliability, contracts and price, none of which he independently compares here.

Asked how he would respond to a hypothetical acquisition offer for such a business, he says he would seriously consider it, given the risk that competition changes the economics. He credits an early product for making access easier. No transaction or price is confirmed by this discussion.

For investors deploying capital amid uncertainty, he recommends looking for founders deeply committed to a problem and for businesses that could have meaningful impact. He names companies he admires and discloses investments in some examples. Those endorsements reflect his personal exposure and judgment.

Stebbings asks whether faster company growth means faster venture exits. The speakers discuss a reported Cursor transaction, and Murdock speculates about the parties' motives and whether an offer made sense. Neither the transaction terms nor private decision-making is independently established in the interview.

Murdock recalls Flipboard as a company where, in his view, a past acquisition opportunity might have been worth taking. He attributes decisions to his recollection of board discussions and advice at the time. The episode does not verify proposed offers or the motivations of the people involved.

His Flipboard story illustrates a difficult board question: whether to sell when a buyer is interested or continue building independently. He believes the company missed a favorable opportunity, but hindsight does not establish what other outcomes were possible at the time. His account of the offers remains unverified.

Stebbings worries that only very large companies can access public markets. Murdock says opportunities depend on the company and market conditions rather than a single revenue threshold. His comments about what a particular firm could have done are hypothetical and do not establish that an IPO would have succeeded.

He says established software companies need a thoughtful response to AI as users adopt more agentic workflows. He is concerned that adding a superficial assistant may not be enough to protect a business. That is his competitive assessment, and the pace and direction of adoption are uncertain.

He calls the present moment an early phase of more agentic work and says incumbent software companies still have time to develop useful responses. His concern is that a token assistant attached to an old product may not be a durable strategy. This is a forecast about competition, not an assessment of every current product.

On private equity, Murdock says highly leveraged software businesses could face pressure if operating earnings deteriorate during a financial disruption. Firms have time to adjust while financing remains available, in his view. The discussion does not evaluate the debt covenants or solvency of any named portfolio company.

He stresses the interaction between business performance and credit markets: leverage can amplify a decline in asset values or make refinancing harder. The speakers discuss examples of leverage, but no single ratio describes all private-equity investments. His concern is a scenario, not a prediction of a specific default.

The pair discuss a proposal involving potential government participation in a frontier AI business. Murdock questions whether government ownership is necessary and says private-public collaboration can take many forms. His interpretation of political motives is opinion; the conversation does not establish a completed transaction or its terms.

On chip export controls, Murdock says policymakers should articulate and debate a broader strategy rather than adopting isolated rules without context. This is his policy preference. The interview does not analyze the applicable regulations or establish how any company complies with them.

Stebbings asks about security concerns involving open models developed in China. Rather than presenting evidence of a specific backdoor, Murdock predicts that current models will eventually be replaced by new architectures. That prediction does not establish that any existing model is safe, unsafe or free of vulnerabilities.

Murdock expects models capable of more continuous or lifelong learning to change the architecture and training landscape. He describes this as a goal for developers and speculates it could arrive within several years or take longer. The conversation does not prove such a system already exists or will arrive on his preferred timeline.

In his description, continual learning would allow systems to incorporate experience over time instead of relying only on a static trained model. He expects future approaches to differ architecturally from today's systems. The terminology covers active research, and he does not give a demonstrated specification or validated timeline.

He says enterprise data changes over time, and different organizations can use similar types of data in different ways. He sees potential value in data providers as businesses customize systems, while acknowledging that data needs and the market could evolve. Stebbings discloses an investment in a data company when making his own market-size argument.

Stebbings suggests a data supplier could move from selling primarily to frontier labs toward serving enterprises that need supplemental data for specialized models. Murdock sees a potentially large opportunity but provides no evidence for a specific market valuation. The host has disclosed his investment in a company in that field.

Stebbings asks whether the conversation overestimates what can happen in a year and underestimates a decade. Murdock admits that continuous learning could take two or three years, or ten: he does not know. He uses medical progress as an analogy for difficult problems and does not claim that cancer has been solved.

In his view, a combination of financial disruption and slower-than-expected technical progress could disappoint inflated expectations. Conversely, further breakthroughs could support continued development. These are conditional scenarios, and the interview offers no reliable timing or probability for either.

In a quick-fire round, Murdock speculates about possible listings and future Nvidia valuation. He also says Meta, Google and Microsoft have large customer bases that he believes may give them time to adapt, while acknowledging differences among them. These are investment opinions, not confirmed IPO plans, price targets or financial advice.

Asked about Apple's AI approach and private-equity firms, he says he lacks enough information to assess the former decisively or rank every firm's portfolio. He discloses his connection to Insight Partners and says its own private-equity exposure is limited. His comments on relative risk are not an independent audit of any fund.

He praises several venture firms for their positions in AI, including firms with different portfolios. These are subjective assessments and reflect his perspective as an investor. He does not establish future fund performance or suggest that a past investment guarantees future returns.

For a final prediction, Murdock expects blockchain infrastructure to find more use in areas such as agent payments or purchasing inference. He discusses industry security concerns and possible products, but his adoption and technical-security claims are speculative. Stebbings thanks him and closes the interview.