Rosa Del Mar

Issue 37 2026-02-06

Rosa Del Mar

Daily Brief

Issue 37 2026-02-06

Why-This-Ai-Wave-May-Create-New-Companies-And-Where-Value-Might-Accrue

Issue 37 Edition 2026-02-06 6 min read
General
Sources: 1 • Confidence: Medium • Updated: 2026-02-06 16:59

Key takeaways

  • Martin Casado predicts many current generative AI applications will look trivial in retrospect, and says the key watch item is whether early behaviors translate into durable enterprise and consumer workflows over time.
  • Martin Casado disputes that today's AI infrastructure build-out is funded like dot-com, arguing the primary builders now have strong balance sheets and cash flows rather than extreme leverage.
  • Martin Casado argues the 'AI must grow 40x' framing is often misinterpreted because the relevant shift is reallocating spend within existing large businesses rather than requiring proportional total revenue growth.
  • Most current AI spending is going into data-center capacity (GPUs, real estate, power, cooling) rather than mainly into software teams.
  • Consultants estimate that the current AI infrastructure build-out would require AI revenue to grow roughly 40x by 2030 to justify the spending.

Sections

Why-This-Ai-Wave-May-Create-New-Companies-And-Where-Value-Might-Accrue

Generative AI is characterized as a behavior shift plus a large performance jump versus earlier incremental, uneconomic AI waves, used to justify expectations of new 'generational' companies. Value creation is argued to extend beyond frontier model labs into a long tail of modality-specific businesses, with defensibility potentially coming from integration and marketplaces. A practical watch focus is whether early, seemingly trivial applications become durable workflows with retention and willingness-to-pay.

  • Martin Casado predicts many current generative AI applications will look trivial in retrospect, and says the key watch item is whether early behaviors translate into durable enterprise and consumer workflows over time.
  • Martin Casado expects many valuable AI companies to exist beyond frontier LLM providers, including a long tail of generative businesses across image, video, speech, and music modalities.
  • Martin Casado argues that long-term defensibility questions are separate from whether AI companies can be profitable today, and he claims profitable, fast-growing AI businesses already exist.
  • Martin Casado says earlier waves of AI produced mostly incremental improvements (around 20% better outcomes) with poor economics, which helps explain why few iconic AI-native companies emerged before generative AI.
  • Martin Casado argues that early 'toy' use cases can foreshadow major future markets.
  • Martin Casado claims generative AI represents a new user behavior and can be orders of magnitude better than prior approaches, creating conditions for new generational companies.

Bubble-Vs-Systemic-Collapse-Framing-And-Analogy-To-Dotcom

The corpus separates valuation corrections from systemic collapse, using dot-com as an example where leverage/fraud concentration (WorldCom) and an infrastructure glut mattered. The AI cycle is argued to be different on funding structure (strong balance sheets vs extreme leverage), and the duration of a prior glut is offered as a rough historical bound, though the duration claim is not substantiated beyond assertion.

  • Martin Casado disputes that today's AI infrastructure build-out is funded like dot-com, arguing the primary builders now have strong balance sheets and cash flows rather than extreme leverage.
  • Martin Casado argues that the late-1990s dot-com bubble was marked by widespread speculative behavior that he does not observe today.
  • Martin Casado argues that a speculative valuation bubble is different from a systemic collapse and that people often conflate the two.
  • Martin Casado attributes the dot-com collapse largely to a fiber glut financed through WorldCom, compounded by 9/11.
  • Martin Casado claims the fiber glut after dot-com only lasted about four years.
  • Martin Casado says he sees zero indication that an AI-related speculative bubble would produce systemic collapse, citing prior overvaluation episodes in mobile, cloud, and SaaS that did not trigger systemic failure.

Roi-Hurdles-And-Justification-Math

Two separate ROI framings are introduced (a large growth multiple and a $2T annual revenue target) as implied justification thresholds for the infrastructure wave. A key interpretive dispute is whether the growth framing should be read as net-new market expansion versus reallocation within existing large businesses.

  • Martin Casado argues the 'AI must grow 40x' framing is often misinterpreted because the relevant shift is reallocating spend within existing large businesses rather than requiring proportional total revenue growth.
  • Consultants estimate that the current AI infrastructure build-out would require AI revenue to grow roughly 40x by 2030 to justify the spending.
  • An estimate cited in the discussion is that the current AI infrastructure wave would need about $2T in annual AI revenue by 2030 to be justified.

Capex-Cycle-Infrastructure-Not-Software

The spend is characterized as infrastructure-heavy (GPUs, power, cooling, real estate) with multi-year lead times, which supports a planning model where public messaging can diverge from long-horizon capacity commitments. The risk assessment is framed around long-term demand absorption and funder resilience if payback is delayed.

  • Most current AI spending is going into data-center capacity (GPUs, real estate, power, cooling) rather than mainly into software teams.
  • Martin Casado frames the key risk question as whether AI infrastructure is overbuilt relative to long-term demand and whether funders have reserves to prevent systemic unraveling if returns arrive late.
  • Martin Casado argues that CEOs may publicly temper AI expectations while simultaneously planning operationally three to five years ahead because data-center build-outs take years.

Watchlist

  • Martin Casado predicts many current generative AI applications will look trivial in retrospect, and says the key watch item is whether early behaviors translate into durable enterprise and consumer workflows over time.

Unknowns

  • What is the actual trajectory of AI-specific revenue (and/or monetized usage proxies) through 2030 relative to the implied justification hurdles discussed?
  • How much of AI infrastructure spending is incremental versus reallocated from prior compute/software spending within the same large businesses?
  • What is the true financing structure of the AI infrastructure build-out across hyperscalers and their supply chains (including any leverage, vendor financing, or project finance)?
  • What are the lead times and binding constraints for data-center capacity expansion in practice (power availability, cooling, real estate), and how do these timelines align with expected demand growth?
  • Do early generative AI behaviors and 'toy' use cases convert into durable workflows with sustained retention and willingness-to-pay in enterprise and consumer contexts?

Investor overlay

Read-throughs

  • Near term value capture may skew toward AI infrastructure and data center supply chain rather than application software teams, because current spend is described as GPUs, real estate, power, and cooling with multi year lead times.
  • If enterprise AI becomes spend reallocation inside large businesses, the revenue justification burden could be met without proportional net new market expansion, changing how ROI hurdles are interpreted versus a pure 40x growth narrative.
  • Durable value may accrue to companies that turn early generative AI behaviors into sustained workflows, with defensibility coming from integration and marketplaces across a long tail of modality specific businesses.

What would confirm

  • Evidence that early generative AI use cases convert into durable enterprise and consumer workflows, shown by sustained retention and willingness to pay over time rather than novelty usage.
  • Disclosures or datapoints indicating AI infrastructure demand absorption tracks long horizon capacity commitments, including continued utilization and buildout cadence consistent with multi year lead times.
  • Signals that AI revenue or monetized usage proxies rise meaningfully through 2030 and align with implied justification thresholds, whether via new revenue or reallocated budgets inside incumbents.

What would kill

  • AI application usage remains toy like and fails to become durable workflows, with weak retention or limited willingness to pay, undermining the case for long term demand absorption.
  • A data center capacity glut emerges that persists, with underutilized GPU capacity or slowed commitments inconsistent with prior multi year planning assumptions.
  • Financing stress appears across the infrastructure buildout and supply chain, such as reliance on leverage or fragile funding structures that contradict the strong balance sheet framing.

Sources