Artificial Intelligence Model Scaling and Computational Infrastructure Requirements

Executive Summary

Three converging physical constraints, semiconductor packaging bottlenecks, a structural grid power shortfall, and saturated data center capacity, are now the primary determinants of who can scale frontier AI, overriding algorithmic talent and capital availability as the decisive competitive variable. Gartner reported in June 2026 that global data center power demand rose 27% to 132 gigawatts in 2026, and Goldman Sachs Research projects the U.S. faces a 9.3 GW structural shortfall that widens to 45 GW by 2028. These constraints compound one another: an organization that secures chips but cannot access grid power cannot deploy; one that secures power but cannot get advanced packaging cannot train at scale.

• Supply-chain/operations teams: Audit GPU procurement pipelines for dependency on TSMC CoWoS packaging and Ajinomoto ABF substrate supply, both confirmed chokepoints through at least 2027; multi-source now or face stranded project risk.
• Risk officers/investors: Treat power-access rights, not chip ownership, as the scarcest and most durable asset in AI infrastructure; weight investment theses toward actors with contracted grid capacity or on-site generation.
• AI product leaders: Organizations without long-term supply agreements for HBM and advanced packaging will find model-size ambitions capped by physical input constraints through at least 2028, reshaping competitive timelines.

The organizations most likely to maintain frontier AI positions through 2028 are those that secured power, packaging, and memory commitments before the current constraint cycle peaked, not those with the largest capital budgets available today.

The Packaging Layer No One Budgets For

The public narrative about AI hardware constraints concentrates on GPU shortages and TSMC fab capacity. That frame is now outdated. As the AI chip supply chain analysis published by Semiengineering and summarized by IndexBox in July 2026 documents, the true binding constraint shifted downstream from silicon fabrication to advanced packaging, specifically CoWoS substrate integration, by 2025. TSMC identified this shift itself, acknowledging that wafer production was no longer the primary bottleneck.

What makes this analytically significant is the second-order constraint below packaging: the ABF substrate, a resin-based material used to connect GPU dies to HBM memory stacks. Ajinomoto of Japan controls over 95% of global refined ABF supply, a concentration ratio comparable to China's grip on gallium refining noted by the IEA. The 30% price increase Ajinomoto imposed in 2026, combined with a projected 2027 supply gap exceeding 20%, means that even if TSMC expands CoWoS capacity on schedule, the substrate layer could become the next sequential bottleneck. What is not being reported in most enterprise technology coverage is that this substrate dependency, sitting three tiers below the GPU in the supply chain, is poorly visible to procurement teams whose risk models stop at tier-one suppliers.

This supply constraint translates directly into competitive positioning. TSMC prioritizes its largest customers, Nvidia, AMD, and Broadcom, per IndexBox reporting. Organizations that are not in that tier-one preference list face lead times of 36 to 52 weeks for data center GPUs, per Spheron Network's April 2026 analysis. Independent cloud providers have partially absorbed this by aggregating spot capacity, but the HBM shortage identified by Micron Technology as persisting beyond 2026 means that spot availability itself is bounded by the same upstream scarcity. The supply chain is not a single bottleneck; it is a stack of interdependent scarcities, and a delay at any layer strands the entire project.

The geopolitical dimension of this packaging concentration is material. The Brookings Institution's June 2026 national security analysis of overseas AI data centers draws an explicit comparison to semiconductor supply chain concentration in Taiwan, warning that AI training clusters that are costly and slow to reconstitute could generate a similar strategic dependency if geographically concentrated. These technology and geopolitical risk dimensions are mutually reinforcing: a Taiwan contingency that disrupts TSMC packaging capacity would simultaneously remove the primary supplier for US hyperscalers and deny Chinese AI developers access to the next-generation packaging that their US counterparts already cannot get enough of.

Power As The New Silicon: Why Grid Access Rewrites The Competitive Map

The power constraint operates through a mechanism distinct from semiconductor scarcity. Chips can be stockpiled, warehoused, and shipped across borders. Grid interconnection rights, once acquired, are durable and non-tradeable in the same way. Gartner's June 2026 analysis makes the competitive implication explicit: "data center power security" is now "the new battleground for scaling and protecting margins in the global AI race." This is not a forecast. As of August 2026, it describes current conditions.

The numbers behind this assessment are striking. Global data center electricity demand reached approximately 485 TWh in 2025, a 17% year-over-year increase, with AI-specific data centers growing at 50%, per the IEA's April 2026 Key Questions report. Gartner's June 2026 analysis projects the demand reaches 290 GW by 2030, from 132 GW in 2026. Goldman Sachs Research estimates the US faces a structural shortfall of 9.3 GW in 2026 that widens to 45 GW by 2028, equivalent to the annual electricity needs of roughly 34 million US households. The IEA notes that data center construction can be completed in two to three years, while energy system infrastructure requires significantly longer lead times for planning, permitting, and construction. This asymmetry between the pace of demand and the pace of supply drives the structural shortfall.

The infrastructure and geopolitical implications compound each other here. The IEA documents that emerging and developing economies account for 50% of global internet users but less than 10% of global data center capacity, and that reliable power is the primary differentiator for which countries can attract data center investment. This spills into sovereignty politics: the EU launched its AI Gigafactories initiative on July 30, 2026, specifically targeting EUR 30 billion in investment partly to prevent European AI development from depending on US or Asian data center infrastructure. The Bloom Energy 2026 Data Center Power Report documents widening grid interconnection timelines, noting that AI-driven compute demand has been outpacing the grid's ability to deliver power since 2025, with pressures intensifying rather than easing into 2026.

Short-term gain, long-term cost: hyperscalers securing power purchase agreements with utilities now, including unprecedented commitments to nuclear and SMR developers, are locking in grid access at current negotiated rates. The IEA's April 2026 report notes that the SMR pipeline with data center operators grew from 25 GW at end-2024 to 45 GW by April 2026, an 80% increase in 16 months. Organizations that delay these commitments will pay a premium at renegotiation, or find that viable sites have been contracted by earlier movers, extending their capacity timelines by several years.

Who Can Still Compete: The Fragmented Competitive Landscape

Not all AI development organizations face the same constraint profile. The competitive landscape is fracturing along three axes: access to advanced packaging slots, access to contracted grid power, and access to capital at the scale required to negotiate both. The hyperscalers, Alphabet, Amazon, Meta, Microsoft, and to a lesser extent Apple, sit at the intersection of all three advantages. Axis Intelligence's compilation from Q1 2026 earnings guidance shows these five are collectively directing approximately $725 billion toward AI infrastructure in 2026 alone. This scale of commitment generates preferential supplier relationships that are structurally unavailable to organizations spending at lower magnitudes.

The second tier consists of frontier AI labs, Anthropic, OpenAI, xAI, and their international equivalents, which are not building data center infrastructure at hyperscaler scale but are securing access through equity relationships, cloud credits, and long-term partnership agreements. AMD's July 2026 acquisition of inference optimization firm Taalas, and its equity-linked Anthropic commitment, illustrate the mechanism: silicon suppliers are moving toward model developers, and model developers toward silicon, to lock in mutual access in an environment where open-market procurement is unreliable. The 247WallSt August 2026 analysis of AMD notes that non-China demand from Anthropic, Meta, OpenAI, and Microsoft now dwarfs AMD's China export-control exposure, illustrating how supply agreements between labs and chip vendors are reshaping the competitive landscape below the hyperscaler level.

The third tier, enterprise organizations attempting to build or operate proprietary AI infrastructure, faces the most acute constraint. World Wide Technology's March 2026 analysis finds that most enterprise data centers are hitting power ceilings, cooling limits, and rack density constraints after committing to GPU hardware, discovering infrastructure limits after the hardware arrives. The Deloitte 2026 State of AI in the Enterprise report identifies legacy data and infrastructure architectures as a primary barrier, noting that systems designed for conventional workloads cannot support real-time AI at the loads frontier models require.

The EU's competitive position warrants specific attention as a coalition fracture point. The EU is not a unitary actor on AI infrastructure investment. Member state industrial capacity, energy mix, and regulatory posture differ materially. The EU AI Office's July 2026 frontier AI findings on EU competitiveness, sovereignty, and security acknowledge structural gaps that the 2026 State of the Digital Decade report also flags, noting progress toward 2030 goals but urging action on structural gaps. Individual member states with favorable power conditions, including France's nuclear base and the Nordic countries' renewable surplus, are better positioned to attract compute investment than coal-dependent eastern European economies. The EUR 30 billion AI Gigafactories initiative announced July 30, 2026, by the European Commission aims to pool capacity, but funding disbursement and site-selection politics across 27 member states introduce execution uncertainty that a national program would not face.

The following table presents the assumptions underlying this assessment, the evidence supporting each, the evidence that would falsify it, the impact on the assessment if the assumption proves wrong, and the single best observable metric for each.

Counterarguments

1. The efficiency improvement argument undermines the constraint severity: The IEA itself presents a High Efficiency scenario in which stronger progress on software and hardware efficiency reduces global data center electricity demand to approximately 970 TWh by 2035, roughly 15% below the base case, with 2.6% of global electricity going to data centers. The Stanford HAI 2026 AI Index documents what researchers call the "jagged frontier" of AI capabilities, noting that efficiency breakthroughs can offset demand growth in ways that linear projections miss. If techniques like distillation, quantization, and inference optimization continue improving at current rates, the conclusion that physical infrastructure access determines competitive advantage could be overstated. The picture is mixed: the IEA's Headwinds scenario shows demand plateauing at around 700 TWh, roughly 40% below the base case trajectory. Organizations betting solely on infrastructure moats could be outflanked by competitors achieving equivalent capability at lower compute cost.

2. The hyperscaler spending narrative underweights the role of open-source and distributed compute: Public discourse on AI infrastructure concentrates on the largest named actors, creating a selection bias toward the most visible constraint patterns. The EU AI Office selected the EUROPA consortium in June 2026 to build a European open-source frontier AI model across 24 EU languages, deliberately circumventing the hyperscaler dependency model. DeepSeek's January 2025 release of R1, which the Stanford HAI 2026 report notes briefly matched the top US model on benchmarks, demonstrated that capability at the frontier does not require proportional compute investment if architectural and training efficiency gains are present. This creates a genuine alternative hypothesis: the constraint analysis here may underweight the possibility that the next capability advance comes from compute-efficient architectures rather than raw infrastructure scale, which would partially decouple competitive advantage from physical infrastructure access.

3. The geopolitical framing may overstate the durability of the US infrastructure lead: The Brookings Institution's June 2026 security analysis acknowledges that the semiconductor analogy to Taiwan is "useful, if imperfect," noting that AI infrastructure is unlikely to reach that degree of concentration. The Stanford HAI 2026 AI Index shows the US leads on top-tier model production and high-impact patents, while China leads on publication volume, citations, and patent output. South Korea leads the world in AI patents per capita. This diversity of capability distribution suggests the competitive landscape is less consolidated than the infrastructure analysis implies. The assessment that hyperscaler dominance is "durable" through 2028 rests on the assumption that no geopolitical, regulatory, or technological shock redistributes supply-chain leverage, which the evidence base cannot rule out with confidence.

Indicators To Watch

The table below presents observable signals that would update the core assessment. Each threshold represents the point at which the assessment's primary findings would require revision.

Near-term watch list: (1) TSMC Q3 2026 earnings call (October 2026) -- CoWoS capacity guidance and any revision to the 80% CAGR trajectory will be the single most informative signal on how quickly the packaging constraint eases; (2) Nvidia and AMD Q3 2026 earnings (October/November 2026) -- any change in supply-driven revenue guidance or HBM allocation language will update the assessment of the constraint timeline; (3) FERC interconnection queue data (Q3 2026 update) -- growth or decline in the backlog will measure whether grid access is improving or worsening ahead of the 2028 Goldman Sachs shortfall projection.

Decision Relevance

Scenario A (~55%): Constraints persist through 2027-2028 with marginal easing, incumbent advantage entrenches. The CoWoS packaging gap narrows slowly, HBM remains undersupplied, and grid interconnection timelines stay long. The organizations that locked in supply agreements, hyperscalers and their closest lab partners, extend their operational capability lead because new entrants cannot access equivalent compute on commercially viable timelines.

If you have an AI product roadmap dependent on scaling model size or inference throughput through 2027, contract for cloud capacity now at multi-year rates, even at a premium; the spot market will not reliably deliver the compute needed for competitive training runs. If you lack that dependency, maintain optionality by monitoring TSMC CoWoS quarterly guidance before committing to on-premise GPU infrastructure that could arrive into a resolved or worsening constraint environment.

Scenario B (~30%): Efficiency breakthroughs decouple capability from infrastructure scale, redistributing competitive advantage. A new architectural or training efficiency advance reduces the compute requirements for frontier capability by 40-60%, analogous to the efficiency gains documented around DeepSeek R1 in early 2025. Infrastructure moats depreciate faster than the current trajectory implies.

If you are a risk officer allocating between infrastructure-heavy and model-efficiency-heavy AI investments, weight the latter more heavily than the current market narrative suggests; the IEA's High Efficiency scenario is not a tail risk but a plausible base case if hardware efficiency trends documented by Stanford HAI continue. If you are a hyperscaler, this scenario represents the primary strategic threat to capital-intensive infrastructure lock-in.

Scenario C (~15%): Geopolitical shock, Taiwan contingency or major supply-chain disruption, triggers emergency reallocation of global AI infrastructure investment. A disruption to TSMC operations, whether through conflict, natural disaster, or regulatory intervention, removes the foundry that the Stanford HAI 2026 AI Index confirms fabricates "almost every leading AI chip." This is not a high-probability scenario, but the Brookings Institution flags it explicitly as the closest analogue to the strategic dependency that already defines the semiconductor industry.

If you advise on technology supply-chain policy or hold positions in defense-adjacent AI infrastructure, this scenario should already be part of your contingency planning; the EU Gigafactories initiative and TSMC's Arizona expansion exist precisely because governments have concluded the tail risk is not acceptable. If you are an enterprise technology buyer, the practical implication is to reduce single-source dependency on any tier-one AI hardware supplier and maintain at least one qualified alternative in each critical component category.

Expert Integration

Expert Consensus Assessment

Analysts from the IEA, Gartner, Brookings Institution, Stanford HAI, and Goldman Sachs broadly agree that physical infrastructure constraints, power and packaging specifically, are the binding variables for near-term AI scaling. There is less consensus on whether efficiency improvements will offset those constraints before they produce durable competitive stratification.

Expert Disagreement Areas

• Severity and duration of the power constraint: The IEA presents a range from a Headwinds scenario plateauing at roughly 700 TWh to a Lift-Off scenario substantially exceeding the base case, reflecting genuine uncertainty about both AI adoption rates and grid buildout pace. Gartner's June 2026 estimate of 290 GW by 2030 sits in the middle of this range but carries its own uncertainty bands.
• Whether efficiency gains can offset scaling constraints: The IEA High Efficiency scenario and the DeepSeek R1 precedent (Stanford HAI 2026) suggest constraint severity may be overstated if algorithmic efficiency continues improving; Micron Technology's persistent HBM shortage statement suggests the picture is not resolving quickly.
• Competitive implications for non-US actors: Brookings explicitly cautions against over-generalizing the Taiwan semiconductor analogy to AI infrastructure concentration, while the IEA documents the structural access gap between developed and developing economies as a near-certain outcome of current investment patterns.

Systematic-Expert Alignment

Alignment: MIXED

This assessment aligns with expert consensus that power and packaging are the primary near-term binding constraints. It diverges slightly from the more optimistic efficiency-improvement framing by weighting the duration of the constraint cycle higher, given Micron's explicit statement that HBM shortage will persist beyond 2026 and Ajinomoto's 2027 supply gap projection. The 55% probability on Scenario A reflects that the evidence base leans toward persistence of constraints but cannot rule out the efficiency-breakthrough alternative that several credible sources keep open.

• Precise data on hyperscaler-specific supply agreements, including terms, volume, and duration of chip and power purchase contracts, is not publicly available. This assessment infers preferential access from spending levels and disclosed equity relationships, not from contract visibility.
• The ABF substrate supply gap projection of 20%+ for 2027 rests primarily on Ajinomoto's own forward guidance, which carries inherent bias; independent corroboration of this figure from buyers or competing analysts is limited in available sources.
• China's domestic AI infrastructure buildout, including indigenous advanced packaging capacity at SMIC and state-directed memory investment, is poorly documented in English-language sources. The finding that Chinese AI developers face more acute packaging constraints may underestimate domestic workarounds not visible to Western analysis.
• Energy demand projections from the IEA, Gartner, and Goldman Sachs diverge materially across scenarios, and the IEA explicitly acknowledges substantial uncertainty about current data center consumption levels, not only future projections. Point estimates embedded in this analysis should be read as scenario midpoints, not confident forecasts.
• The geopolitical risk to TSMC operations, while flagged by Brookings and Stanford HAI, has not materially disrupted supply as of August 2026. Treating Taiwan contingency as a tail risk rather than a base case is the appropriate calibration, but the assessment could underestimate how rapidly the constraint profile would change if that risk materialized.