Rising bond yields have put the time value of money back on the table. Investors are right to watch borrowing costs, distant cash flows, and businesses that swallow capital by the billion. Across much of innovation, however, the price of taking too long deserves equal attention. We calculate the cost of capital to two decimal places. The cost of the calendar often receives barely a sentence.
A valuation model can make waiting look deceptively harmless: move the revenues six months to the right, apply the discount rate, and carry on. In our Super-Moore era, those six months can change the business itself. Customers sign elsewhere, competing products improve, and expensive hardware loses part of its commercial life before it earns its first dollar. A company may secure better financing terms and discover that it has financed a worse opportunity.
The delays that produce this outcome are usually quite ordinary. A founder holds out for a better valuation. A sponsor is traveling. Management wants a consultant to bless what its operators already know. Each request sounds reasonable, and the people involved often believe they are protecting the business. Yet the savings from another negotiation are easier to demonstrate than the value lost while everyone waits. The committee records the discount it secured; the customer who moved on never appears in the minutes.
Our management habits reinforce the problem. We learned to keep urgency and long-term planning in different rooms. An immediate problem demanded action; a five-year ambition earned an off-site. Unfortunately, in the present day, even preparations for a technology, a model, or a plant that will be cutting-edge years later need to be handled with the urgency of a house on fire.
Last week offered a small and tragically comic proof. The world decided again that AI needed rules, this time with more fervor than ever. Articles went around like a cold. By Monday, the feed had moved on, with responsibility transferred to amorphous policymakers by everyone who wrote and spoke. Well, if you take the darkest view, as implied by the loudest voices, hidden agent systems thicken while those rooms schedule their next session. If agent swarms can hide and coordinate through tens of thousands of messages today, aren’t they already learning how to bypass any controls we can think of?
We do not share the darkest predictions, but the mismatch is plain enough. Companies, regulators, and investors can no longer afford to spend months preparing to address a world that changes during the preparation. Across semiconductors, energy, medicine, and AI, this widening gap between the pace of innovation and the pace of decisions is reshaping investment returns. It rewards businesses that help others move sooner and raises the cost of familiar organizational habits or bureaucratic delays. It also makes a prompt, correctable decision potentially more valuable than a polished decision delivered too late. The following examples examine how much can change while a perfectly sensible proposal waits for its next meeting.
Two Identical Buildings: The Arithmetic of the Lag
Consider two hypothetical data centers selling AI computing services. They sit in the same geography, house identical liquid-cooled accelerator clusters, connect to identical network fabrics, and draw from the same power envelope.
Building A flipped the breakers in January. Building B, hobbled by permitting delays, an unaligned joint venture, and internal design reviews, came online nine months later in October.
A simplistic valuation model treats that gap as a straightforward arithmetic deferral: Building B captures Building A’s revenue run-rate, just three quarters later. In practice, those nine months can change the economics of the entire investment.
The price of the output can fall sharply. The cost of accessing a given level of AI capability has been declining rapidly. That puts pressure on what computing providers can charge, although the effect depends on their contracts, workloads, and competitive position. Building B may open into a market offering substantially less for the work its equipment performs. Growing demand could offset some of that pressure, but the original revenue forecast needs to be rebuilt.
The equipment bill can rise while the project waits. If Building B deferred procurement while resolving its approvals, it may find that memory, networking equipment, or complete systems have become harder to obtain. Allocations can tighten, delivery windows can slip, and securing the original configuration may require a premium. If it bought everything early, it faces a different cost: expensive equipment sitting idle before earning its first dollar.
The hardware moves further through its competitive life. Building B’s accelerators are nine months behind the technological frontier when they enter service. They can still perform valuable work, but newer systems may offer customers better performance or lower operating costs. The period in which its equipment commands attractive pricing may already have shortened.
For illustration, suppose Building A’s original business case projected an eighteen-month capital payback. Building B cannot assume the same eighteen months simply begin in October. Lower realized prices, a higher equipment bill, or weaker utilization could substantially stretch its payback, potentially into several years. Meanwhile, Building A has already earned nine months of revenue, recovered part of its investment, and established customer relationships. Building B may have to compete aggressively on price while carrying a heavier unrecovered investment.
The financial community examines these projects in extraordinary detail. Investors scrutinize take-or-pay contracts, customer creditworthiness, vendor guarantees, financing terms, and residual asset values. All deserve attention. But their value depends heavily on when the facility can begin delivering what it has promised. A contracted customer still needs capacity on schedule. A vendor guarantee may offer compensation without restoring the months lost.
The delays can originate almost anywhere. Government permissions take longer than expected. A power connection slips. A smaller operator lacks the purchasing power to secure equipment while larger customers move ahead in the queue. These differences can determine which facilities capture customers, generate cash, and finance their next expansion.
Traditional infrastructure has always suffered from delays and cost overruns. Innovation adds another exposure: the technology, selling price, and competitive conditions can change materially before construction is complete. For a heavily indebted operator, a few months can therefore disrupt far more than the timing assumptions in an IRR calculation.
Execution speed deserves a much more prominent place in the investment assessment of hyperscalers, neoclouds, and their supporting infrastructure. Access to competitive hardware, dependable power, and timely approvals must come together. An operator that brings productive capacity online sooner can begin earning, learning, and strengthening its customer relationships while another is still paying to prepare. That head start can support its next investment, making the cost of a missed opening date extend well beyond the first project.
The Model-Maker’s Treadmill: Building for Quarters to Reign for Months
If the arithmetic of data centers looks punishing, the reality facing the frontier model-makers is downright ruthless. Nowhere else in modern capitalism does an enterprise spend twelve to eighteen months marshaling billions of dollars in compute, curating petabytes of synthetic and human tokens, nursing fragile training runs through hundreds of hardware faults, and subjecting the output to exhaustive safety audits, only to produce an asset whose undisputed commercial reign lasts barely six months.
Developing a frontier model remains a massive multi-stage undertaking. The raw pre-training run alone occupies months of uninterrupted, highly synchronized compute. Epoch AI estimates roughly three months of pure cluster execution for architectures on the scale of GPT-4, Llama 3.1’s 405-billion-parameter model, and Grok 3.
Yet pre-training is merely the middle chapter of a far longer development program. Data sourcing, synthetic data generation, architectural validation, reinforcement learning, post-training alignment, evaluations, safety red-teaming, and inference optimization routinely stretch the full cycle to a year or more from initial commitment to public availability.
The asymmetry lies in what happens once the model enters production.
The tenure of undisputed frontier capability has compressed to a matter of months. Claude Opus 4.1 was retired exactly twelve months after its introduction. GPT-4o, which appeared to be an unshakeable consumer fixture for a time, stepped aside as the primary default within eighteen months. Several members of the subsequent generation held their position as the industry benchmark for barely a quarter before an internal increment or an external rival took the chair.
Those initial frontier months matter disproportionately. They represent the narrow harvest window in which superior benchmark performance commands premium pricing, wins enterprise platform evaluations, and persuades developers to anchor their software workflows around the API. Revenue may well endure after a successor arrives, particularly among enterprise customers that prioritize operational stability, known latency, and predictable costs. But the period during which superior capability alone drives commercial sales is fleeting.
If we have to dramatize in the language of yore, they are building an aircraft carrier on a multi-year construction schedule to fight a naval battle that lasts a single weekend. Or, more appropriately, they are built like the most expensive movies, whose success is measured by the revenues of the first few days.
It is worse for the model-makers. Because development expenditure begins several quarters before that window opens, the capital commitment for the subsequent generation must be signed long before anyone knows how much cash the current release will return.
The Fallacy of the Post-Mortem
The fatal error in this environment is attempting to run model development like traditional software engineering.
For three decades, technology leaders were schooled in the gospel of measured validation: ship a product, listen to customer feedback, conduct a thorough retrospective, identify product-market fit, and only then greenlight the budget for the next version.
In frontier AI, that methodical sequence is an act of commercial suicide.
If a research lab waits for Model N to launch, monitors user adoption for a quarter, and then convenes an off-site retreat to decide what Model N+1 should look like, it has already signed its own obsolescence.
By the time they draft the architecture specs, clean the datasets, and reserve the clusters, two full generational cycles will have passed. They will emerge eighteen months later with a model that answers the questions the market was asking two years ago.
The labs that have captured leadership in this cycle understood this paradox early: product development cannot be sequential; it must be concurrent.
When Anthropic delivered the Claude 3 family in early 2024, they did not pause to bask in the critical praise for Opus. The training runs and data regimes for Sonnet 3.5 were already well underway before the public had even finished testing Claude 3.
By the time competitors scrambled to match the original Claude 3 benchmarks, Sonnet 3.5 dropped like a thunderclap in mid-year, completely resetting the industry’s coding and reasoning standard at a fraction of the operating cost.
OpenAI demonstrated the same relentless temporal discipline. While the market was still obsessing over standard transformer scaling and prompt engineering, their research pipeline had already committed massive compute tranches to parallel tracks: architecting next-generation foundational scaling while quietly developing inference-time reasoning architectures long before the broader market realized that thinking time was the next frontier.
These organizations did not wait for validation. They planned three horizons at once:
- Horizon 1 (The Present): Model N is serving enterprise clients, collecting telemetry, and generating the cash flows needed to offset immediate operating expenditure.
- Horizon 2 (The Mid-Flight Run): Model N+1 is already burning megawatts in the training cluster, with its dataset frozen and its hyperparameters locked down months ago.
- Horizon 3 (The Whiteboard): The architectural bets, synthetic reasoning loops, and multi-gigawatt power contracts for Model N+2 are already being negotiated, indifferent to whether Model N is currently receiving rave reviews or lukewarm indifference.
One only needs to look at what happens to the organizations that hesitated. Several well-capitalized institutions, enterprise software giants, and cautious research consortia entered this race with pristine balance sheets and world-class researchers. But their decision-making culture was infected by institutional prudence. And their fall is something we do not need to detail.
By now, this has turned into a perpetual slipstream for the laggards. Without the premium margins of a market leader, their cash flows cannot fund the exponential capex required for the next cluster.
This is true across innovations, including in hardware. Once you fall a single generational beat behind in this race, catching up requires an order higher capital, materially higher risk tolerance, and an opponent willing to pause.
The World of Queues
A new chip can be announced to the whole world on the same day. It will reach customers on very different dates. Some help develop the systems around it, prepare their facilities early, and secure initial deliveries. Others join the queue after the performance figures have persuaded them to buy. By then, the first group may already be offering customers something the second can only describe in a proposal. There is not much new in this: a technology’s existence and a business’s access to it are separate events.
In the world of shortages, the gap takes draconian forms. High-bandwidth memory, advanced packaging, optical interconnects, and power equipment each bring their own constraints. A company can secure most of what it needs and still be held up by the remaining piece. An expensive cluster awaiting its networking equipment has something in common with a plane grounded for a missing washer: you cannot fly with a loose nut.
This helps explain why long-term supply agreements have become so consequential. Investors accustomed to the purchasing power of technology giants can find it puzzling to see them scrambling for components from much smaller suppliers. But a component’s share of the equipment bill says little about the cost of its absence. The supplier holding up the launch acquires a bargaining power that its size alone would never suggest.
The commitments now extend across memory, foundry capacity, advanced packaging and optical components. Power equipment offers another revealing example: GE Vernova has described customers paying substantial deposits to reserve turbine production slots before their construction arrangements are complete. These buyers are committing money early because a place in the schedule has commercial value. By the time every other detail is settled, that place may belong to someone else.
For a customer still negotiating, these commitments change the landscape. Another round of bargaining may improve the quoted price while pushing delivery into a later production window. Missing that window can postpone system testing, customer validation, and the eventual start of revenue. A fortnight spent seeking a discount can travel through the project and emerge as a quarter’s delay. Procurement reports a saving; the business inherits the wait.
Nor does finding an alternative supplier necessarily solve the problem, as Apple seems to have found recently in trying to procure cheaper memory. A replacement component may need qualification, firmware changes, or fresh reliability testing. Manufactured equipment must still be shipped, installed, and validated as part of the complete system. The advantage of scale therefore reaches beyond the familiar ability to negotiate a lower price. Large, credible buyers can commit earlier, support capacity expansion, and work with suppliers before products are ready for broad release. Those relationships give them a better chance of aligning deliveries, engineering support, and deployment.
Long-term commitments carry risks: demand can disappoint, designs can change, and yesterday’s scarce component can become tomorrow’s surplus. The judgment lies in deciding which capacity is important enough to secure before the need becomes obvious. In a market organized around bottlenecks, waiting for certainty can mean accepting whatever remains.
The Order of Arrival
A technological breakthrough can be right about the destination and wrong about how soon customers need to travel there. While its developers resolve manufacturing problems, an established supplier finds another way to stretch the existing design. Customers facing an immediate bottleneck buy what works. The elegant successor eventually arrives to discover that the supposedly temporary solution has acquired customers, production experience, and another round of improvements.
High-bandwidth memory illustrates the problem. Hybrid bonding connects chips directly without the tiny solder bumps used in conventional stacking, offering a route to denser connections. Yet SK hynix announced a 16-layer HBM3E design using its advanced MR-MUF process while continuing to develop hybrid bonding. Improving the established method created another option for taller stacks.
Lithography offers more examples where progress is not about quickly adopting the best. China’s chipmakers have continued to extend deep-ultraviolet lithography through multiple patterning, more complex processing, and changes to chip design. Additional patterning adds processing time, complexity, and yield risk, but it allows development to continue with available equipment.
The lesson travels beyond the companies facing restrictions: an apparent equipment limit can become a process-engineering problem. That does not establish that unconstrained manufacturers are copying China’s methods. It does give everyone a reason to examine how much useful life remains in already installed equipment.
TSMC faces a different choice. It already uses EUV, yet continues improving masks, pattern accuracy, and manufacturing processes while preparing for the more advanced High-NA generation. In September 2026, TSMC and ASML stated that TSMC intends to begin using High-NA in high-volume manufacturing in 2030. Its engineers therefore have several more years in which to extract value from the existing platform. The newest machine has to justify its place in the production schedule.
Optical networking offers a similar contest. Moving optical connections closer to the main chip can reduce power consumption and shorten electrical paths. Meanwhile, pluggable modules keep improving. Broadcom’s Sian3 announcement described support for 1.6-terabit modules with more than a 20% power reduction against its predecessor. That extends the appeal of equipment operators already know how to install and replace.
For investors, this makes the development calendar part of the competitive analysis. A three-year roadmap must anticipate three years of improvement elsewhere. The relevant comparison is between what a company can deliver then and what customers will be able to buy then.
The Next Meeting Is Next Month
Innovation changes its timetable whenever it finds a better answer. Institutions keep the one already circulated. The annual budget, the quarterly committee, and the strategy blessed at last year's off-site all hold their authority long after the assumptions underneath them have expired. The calendar was built to provide continuity. Increasingly, it provides an alibi.
Memory makers allocate next year's output in July and August. A capex plan approved in the October budget round arrives at a counter that closed two months earlier, to be offered 60 to 70 percent of what it asked for, at a higher price, on terms stretching up to five years.
A Japanese container ship began carrying commercial cargo under Level 4 autonomy in January 2026. The global code governing such ships becomes mandatory in 2032. Six years of voyages will be logged before the rules binding them exist.
Hospitals once planned an MRI over years: shielded room, reinforced floor, helium plumbing. A portable scanner now rolls through a standard doorway and plugs into the wall. It does not replace the big machine, yet a good share of a two-year committee debate concerned a room that some scans no longer need.
Standards bodies deserve credit for noticing. From October, ISO and IEC require a new international standard to be completed within 36 months of approval, or 48 with an extension, or it is canceled. Three years is the reform. On a chip roadmap, three years is two generations.
JEDEC published the HBM4 standard in April 2025, setting the data rate at 8 gigabits per second per pin. Within months, Nvidia asked for 10, then 11. All three memory makers had samples running above 11 by the end of that year, and Samsung's commercial HBM4 now runs up to 13. The standard became a floor before most accelerators using it had shipped. For the next generation, HBM4E, vendors are building custom base dies to individual customer specifications with no unified JEDEC standard yet published. The rulebook still exists. The biggest buyer writes its own.
Paradoxically, startups reproduce the same rigidity with fewer forms. Milestones agreed with investors nine months ago fund a capability that is now cheaply available. Changing course needs a board meeting, a revised budget, and consent from diaries already full. The founder is told to be entrepreneurial inside the fence of an old forecast.
Thinking Fast and Thinking Very Fast
Three themes keep returning in our work: innovation is becoming more complex, moving faster and, in many businesses with the strongest competitive defenses, demanding enormous capital. Many investors still long for the elevator pitch, leisurely deliberation and a modest cheque that compounds into a fortune through network effects. Exceptions aside, the current innovation era differs from the previous one by being at the opposite end on each of these axes.
A financial investor can sit this out, perhaps awaiting the satisfaction of saying, “I told you so.” Companies and nations have less freedom. Their competitors keep acquiring customers, securing capacity, and building capabilities. Investing only in what one understands is sensible discipline. However, in the current era, that means an obligation to keep learning, not to wait for things to turn or be made simpler.
Every organization has its brooders, almost always at the decision-making levels. They want the quarter to settle, the data to mature, and the market to stabilize. They complain about the pace as though it were a passing disturbance a patient committee could outlast. A business must increasingly create stability inside its operations while adapting to change outside. Waiting for universal calm can become a permanent occupation.
Long-term planning now requires urgency and the ability to revise commitments as reality unfolds. Management must absorb evidence and correct course while work continues. An organization that needs six months to recover its composure after every change will spend much of its commercial life recovering.
Yet our institutions scrutinize spending far more closely than postponement. A capital request needs forecasts, sensitivities, and signatures. Another month’s delay may require only a senior executive saying, “I’m not sure.” That could reflect insight, an unread paper, or reluctance to decide. In most organizations, nobody asks which.
Errors of commission have names attached. Errors of delay disappear between meetings. The accounts record the expenditure; they rarely identify the customer who signed elsewhere or the allocation window that closed while the memo circulated. Delay can acquire a reputation for depth simply because its costs are harder to attribute.
A proposal to wait should explain what the waiting will improve, what evidence will arrive, and what may be lost in the meantime. “Let’s wait” has the potential to turn tremendously costly in the innovation segments. As long as those deciding to wait are aware of the costs and see other options as far worse, it is perfectly fine.
Financial investors have it relatively easy, but even they need to go beyond opportunity costs to incorporate the costs of delay into their fundamental analysis. Bond yields will keep moving, and another fifty basis points will generate plenty of commentary. But precision about the discount rate offers little protection against months of organizational hesitation.
The bond yield tells us what it costs to borrow the money. The calendar helps determine whether the opportunity will still be there to repay it. In much of innovation, the most expensive item on the term sheet may be the quarter spent getting comfortable enough to sign.
