The AI Path Is Uncertain. Build a Portfolio of Bets.

AI has evolved from a productivity experiment to a strategic force reshaping capital allocation, operating models, and competitive positioning. The first article in this series examined how AI is changing the sources of competitive advantage—and how companies can determine which of their moats are strengthening, eroding, or being hollowed out. Yet the path AI will take to transform competitive advantage may take several turns.

Four uncertainties will shape this AI transition over the next five years: how far the technology advances, what compute will ultimately cost, whether the capital pouring into AI earns an adequate return, and whether access to leading models remains open across borders. Each bears directly on the choices companies make today in planning for the AI transition and how they strengthen their competitive moats in the future.

Faced with that uncertainty, many C-suite executives and boards are tempted to wait for the picture to sharpen. But waiting does not preserve optionality; it often destroys it. Capabilities take time to build, scarce assets become harder to secure, and competitors keep moving.

The better approach amid uncertainty mirrors that of a disciplined investor. Instead of trying to forecast your way to a single winning position, build a portfolio of bets sized to the uncertainty. Some are no-regret moves that create value across almost any future. Others are options: contained investments that preserve the right to move decisively if a particular future materializes. A few are asymmetric wagers, with a bounded downside and the potential upside to reshape the business.

This article applies that portfolio lens to the four AI uncertainties that matter most. None can be predicted in detail. But each can be understood well enough for companies to build strategies that can withstand adverse outcomes—and exploit favorable ones.

How Much Is Left in Scaling?

The most consequential question about AI’s future is not whether the models keep improving—they will. It is how much more capability can be bought through scale, and at what price. For the past several years, the recipe has been simple and astonishingly effective: more data, more parameters, more compute, and larger clusters. It has turned the language model from a clever text generator into a general-purpose system that can write code, read contracts, digest a quarter’s filings, and run a workflow with diminishing human supervision. The open question is how much further that recipe is likely to run.

There are two plausible paths: either the transformer architecture underlying today’s leading models still has substantial headroom, in which case the next generation of models buys real gains in reasoning, reliability, and autonomy; or the industry is entering a phase of diminishing returns to scaling, in which each increment costs more, delivers less, and matters less to the average enterprise. No one knows which path will prevail, and the people closest to the frontier disagree.

But here is the point most boards miss as they wait for that argument to settle: for most companies, technology has already stopped being the binding constraint. The models on the market today are powerful enough to transform much of a large company’s work, whether or not the frontier advances substantially further. What holds companies back is not the capability of the technology but their own capacity to absorb it—the adoption, integration, governance, and process redesign that turn a capable model into a changed business. The bottleneck has moved inside the enterprise.

The first move is no-regret: build the organization’s capacity to absorb AI, because that capacity pays off regardless of how the technology frontier evolves. This is less about technology than the organizational plumbing around it: evaluation systems that test models against real workflows, governance that lets a unit deploy without a six-month legal detour, and process redesign that turns capability into a changed operation. A company that benchmarks each release against its own internal criteria learns faster than one waiting for the market to deliver a definitive verdict.

The second move buys optionality cheaply: keep the technology stack model-agnostic and experiment across models, including open-weight ones, so the company never becomes captive to a single vendor. Running open-weight models alongside frontier APIs is low-cost insurance. It reveals where a smaller, self-hosted system can already match the performance of a costly frontier model, and preserves an option if economics and control favor models that the organization can run itself.

The third move is a real bet: fine-tune a specialized model in the domain the company understands better than any frontier lab. The wager is that if returns to scale diminish, advantage will shift toward companies that deploy smaller, purpose-built models effectively. A system tuned on proprietary data becomes an edge rivals cannot replicate cheaply. This is a bet, not a hedge: it pays off decisively if scaling flattens, but risks being leapfrogged if frontier models continue to improve rapidly.

Is Cheap Compute a Safe Bet?

Most corporate AI business cases rest on a single, largely unexamined assumption: that compute will inevitably become cheaper. The logic is borrowed from earlier technology cycles—most often the dot-com crash, when bandwidth costs fell over the following decade by an order of magnitude, which helped unlock the consumer internet. Applied to AI, the argument is that inference costs will decline as scaling efficiencies compound, open-source models approach parity with closed ones, and hyperscalers compete away their margins. It is a comfortable analogy, but it may be the wrong one.

AI’s binding constraints are physical rather than a matter of pure silicon scaling, and in the near term, supply shortages are only getting worse. Advanced packaging supply remains short of demand despite TSMC lifting capital expenditure (capex) by roughly 30%. High-bandwidth memory is sold out across all three major producers, with warnings of shortages extending into 2027–2028. Grid-connected power has become the tightest bottleneck of all, with the three at-scale gas-turbine makers booked through 2029 at prices well above year-end 2025 levels.

Compounding this physical scarcity is a financial dynamic that few adopters have fully absorbed: compute today is almost certainly priced below its long-run cost of production, subsidized by venture capital, hyperscaler credit programs, and land-grab consumer pricing. That subsidy is now visibly unwinding. Microsoft Copilot has shifted to consumption-based pricing, and the major labs have raised prices for their higher tiers. Each move is small; together, they signal the direction of travel.

The implication for strategy is uncomfortable but clear. Near-term costs are likely to rise as bottlenecks tighten and below-cost pricing unwinds. If costs do eventually fall, it will likely result from a genuine efficiency breakthrough, a stall in demand, or a capacity glut following overbuilding.

The no-regret move is to manage compute as a scarce, structurally expensive input—not a commodity destined to become free. That means investing in proprietary data that reduces the need for fine-tuning; using orchestration to route simple tasks to lower-cost models while reserving frontier models for genuinely complex work; and prioritizing high-return use cases rather than proliferating pilots. Importantly, there should be a clear link between AI decision making and the company’s existing investment processes—evaluating and prioritizing investments in AI on the same basis as any other investments requiring capital or resources. This discipline lowers the cost of AI deployment regardless of whether compute costs rise or fall.

The option is to gain low-cost protection against sustained scarcity: preserve access to alternative providers and smaller, more efficient models without committing to one. A continuous evaluation program identifies workloads that an open-weight or distilled model can handle at lower cost, while multiprovider access caps exposure if a vendor's pricing or capacity tightens. For a modest investment, the company can shift its compute mix as price or availability change. Developer usage already points in this direction, with open-source models capturing a growing share of tokens routed through OpenRouter, an AI model marketplace and gateway. (See Exhibit 1.)

The real bet is to lock in a scarce input before rivals do. For a nontech enterprise, that means a long-term contract for reserved capacity or committed power at today’s terms, ahead of the queue. The company is betting that scarcity will persist. Once order books fill, competitors that waited will have to pay more or wait even longer. The bet is bounded but asymmetric: the contract will look expensive if a glut drives costs down, but pay off decisively if bottlenecks stay tight.

Can the Revenue Catch the Capex?

The five largest hyperscalers—Microsoft, Alphabet, Amazon, Meta, and Oracle—are on track to commit roughly $5 trillion of AI capital expenditure by 2031. (See Exhibit 2.) The scale of that investment matters far beyond the technology sector. It assumes that AI revenues will rise fast enough to cover not only depreciation but also an adequate return on a capital base measured in the trillions.

That is a demanding proposition. Combined AI revenue today remains only a fraction of what would be required to justify the investment now being made. And labor substitution alone is unlikely to close the gap. The economics ultimately depend on AI creating new sources of value, not simply reducing the cost of existing work. (See “The Revenue Hurdle Behind the AI Capex Boom.”)

Whether those revenues arrive on the timeline that markets are pricing is the central financial uncertainty of the cycle. If they do, the investment will help sustain rapid advances in AI capability and infrastructure. If they do not, the result could be a sharp repricing of the financial structure around AI without making the technology itself any less capable.

Such a correction would reach well beyond technology. With the five largest hyperscalers accounting for nearly 20% of the S&P 500, the effects could transmit through capital markets, consumer demand, and future investment in compute. Risk premiums could rise, equity issuance could close, and debt could remain available mainly to investment-grade borrowers. Falling household wealth could weaken confidence just as hyperscalers cut AI capex. The effect on compute costs would be less clear: overbuilding could make capacity cheaper, while a sharp pullback in investment could constrain future supply.

The no-regret move is to build balance-sheet resilience that pays off regardless of whether the correction comes: extend debt maturities and secure financing while credit remains readily available. A strong balance sheet funds growth if revenue catches capex and is the precondition for acting if it does not. Alongside it, accelerate the unglamorous work in data, workflow redesign, and AI fluency, which compounds regardless of what markets do.

The option is to preserve, at modest cost, the ability to gain advantage if a valuation reset occurs: build the acquisition pipeline now, without committing capital. That means maintaining an up-to-date list of valuable AI assets, talent, and IP; developing relationships with the relevant teams; and securing advanced funding capacity, deal resourcing, and alignment on go/no-go approval criteria. If valuations reset, the company can move in weeks rather than quarters. A small investment and no commitment turn a market dislocation into a buy-list the company is ready to act on.

The real bet is a choice about timing: buy the decisive capability now, or hold dry powder in anticipation of a valuation reset. Buying now secures the asset but requires paying at today's potentially inflated valuation. Waiting risks more than a correction that never comes. The asset may be gone before it arrives, sold to a rival who couldn't wait, or diminished as the talent and momentum that made it decisive dissipate. The wager is that the market corrects while the asset remains both intact and available. Its potential payoff is a strategic position that rivals will struggle to replicate once markets recover.

Will Geopolitics Keep the Playing Field Level?

The assumption that frontier capability will remain available to anyone willing to pay has now been tested. On June 12, 2026, the US Department of Commerce ordered Anthropic to suspend its newest models, Claude Fable 5 and Mythos 5, for all foreign nationals. Because citizenship could not be filtered in real time, the models went offline worldwide within hours. Fable 5 had been publicly available for three days. It was the first US export-control directive to retroactively suspend access to a commercially available AI model.

On June 30, only 18 days after the directive, the Commerce Department lifted the restrictions and access was restored. It did not publicly disclose the basis for lifting the restriction. The reversal did not change the underlying implication: access that appeared commercially available and durable could be withdrawn with almost no warning.

The lesson runs deeper than “frontier access is rationed.” That framing treats the risk as a queue: a company gets the leading model later or on worse terms. The Fable ban demonstrated something different: a capability that a company has already deployed and on which it has built a product can be revoked by a government acting within its own regulatory authority, without prior notice. For any non-US organization, the model underneath the business is held on a license that a foreign government can cancel overnight. This is not procurement risk; it is continuity and concentration risk, on the same register as a single-supplier or single-jurisdiction dependency.

The instinct to pursue a sovereign model is largely impractical: genuine control of the entire AI supply chain lies beyond the reach of any single enterprise. The more realistic objective is execution sovereignty—the ability to run inference on infrastructure you control, within a jurisdiction whose rules you understand and can rely on.

That brings open-weight models—such as Meta’s Llama, Mistral, and Chinese offerings including DeepSeek, Qwen, GLM, and Kimi—squarely onto the board agenda. Once self-hosted, such a model cannot be withdrawn or remotely disabled by its developer, even if its capabilities lag those of the most advanced proprietary models. But that resilience comes with provenance and compliance risks. The use of a Chinese-developed model may be incompatible with the requirements for US government, defense, financial services, or other sensitive work, and approval in one jurisdiction offers no guarantee of acceptance elsewhere.

The no-regret moves are to classify workloads according to the consequences of losing model access; maintain a model-agnostic stack; and never assume access to the leading models will remain equal, durable, or unconditional. For a small investment, this approach pays off in every scenario. It lowers switching costs, reveals which workloads truly require frontier capabilities, and ensures that a sudden withdrawal of access triggers a prepared response rather than a crisis.

The option is to establish a viable fallback before it is needed: a tested path to shift critical workloads across providers, model families, hosting environments, or jurisdictions if access changes. Running a self-hosted open-weight model in parallel—even one that trails the frontier models—validates the contingency plan and reveals which workloads it can support. At a modest cost, this turns a Fable-style withdrawal from a business interruption into a controlled, rehearsed transition.

The real bet is to establish execution sovereignty before it becomes necessary: make a bounded investment in self-hosted capability, locally controlled infrastructure, or a strategic model partnership in a reliable jurisdiction. It wagers that access will fragment further—and that the ability to run inference without exposure to a government-directed loss of access will become a decisive edge for regulated or critical operations. The downside is the cost of the investment together with the provenance and compliance risk of a self-hosted or non-Western model. The upside is operational continuity at precisely the moment less-prepared rivals lose access.

A Portfolio, Not a Position

Set the four uncertainties we’ve just reviewed side by side, and they look like four separate problems: about models, cost, capital, and borders. They are not. Although each requires a different response, they all call for the same discipline: avoid building a strategy whose success depends on uncertainty resolving in your favor. A company that needs scaling to continue, compute to become cheaper, AI capex to pay off on schedule, and the playing field to remain open has not built a resilient strategy. It has stacked four bets on one outcome, creating a single fragile plan.

The reason to spread the bets is asymmetry, and it runs in both directions. Take a market correction. A company that has extended its debt maturities and preserved dry powder loses relatively little if the repricing never comes—the cost of carrying a stronger balance sheet for a year or two. But if the repricing does come, that same readiness allows the company to acquire AI assets, talent, and IP it has been tracking at valuations not available today, even as more leveraged rivals are forced to sell. The position that protects the company in an adverse scenario also equips it to profit from one.

The same logic applies elsewhere. A model-agnostic architecture that protects against a Fable-style loss of access also allows a company to adopt a superior model as soon as it becomes available. Secured power that limits compute costs also enables the company to scale when supply-constrained rivals cannot. The aim is not to predict which path the market will take, but to structure the portfolio so that being wrong costs little—and being right pays twice.

Exhibit 3 translates this logic into a representative set of no-regret moves, options, and asymmetric bets across the four uncertainties. The mix will vary by company. No-regret moves should carry most of the weight because they create value across multiple futures. Options should preserve flexibility where the range of outcomes is widest. Asymmetric bets should remain few, selective, and grounded in the company’s particular sources of advantage. The objective is not diversification for its own sake, but a portfolio in which no single assumption can determine the company’s fate.

The returns of the next decade will not accrue to the companies that predict the path of AI most accurately. None will. They will accrue to companies that position themselves most intelligently for the range of paths AI might take. A board cannot decide to forecast the future correctly. It can decide to build a portfolio that does not depend on doing so.