February 24, 2026
Two recent pieces reach different conclusions about AI infrastructure and employment. Citadel Securities published “The 2026 Global Intelligence Crisis”, which argues that AI displacement risk is overstated and adoption follows familiar S-curves. Citrini Research published “The 2028 Global Intelligence Crisis”, a bear-case scenario in which rapid white-collar displacement reduces consumer spending and then damages real estate and private credit.
Their disagreement turns on two questions: what limits deployment, and whether those limits relax before the economy can absorb displaced workers.
The Citadel Argument: Physical Limits Slow Adoption
Citadel cites AI capital expenditure at 2% of GDP (about $650 billion), an 11% year-over-year increase in software engineering job postings, 4.28% unemployment, and roughly 2,800 planned US data centers. It reads those figures as evidence that investment has accelerated while displacement has not.
The report also argues that displacing white-collar work “would require orders of magnitude more compute intensity than current level utilization” and that marginal compute costs could rise above labor costs as automation expands. Its mechanism is supply-side: physical infrastructure limits the rate of deployment.
That mechanism is plausible, but aggregate compute is an incomplete measure. Over the past 20 years, compute throughput has reportedly improved about 60,000 times while memory bandwidth improved about 100 times. The resulting gap makes data movement a limit on useful throughput even as FLOPS become cheaper. In a memory-bound inference workload, adding arithmetic capacity alone does little.
The memory wall shows up in three dimensions:
Capacity: HBM per GPU has grown from 80 GB in the H100 to 192 GB in the B200, but model sizes and context windows have also grown. The KV-cache ceiling, N_concurrent = (HBM - W_model - overhead) / KV_per_session, makes the tradeoff explicit. Holding the other terms constant, more context permits fewer simultaneous sessions per GPU.
Bandwidth: HBM3 delivers 819 GB/s per stack, HBM3e reaches 1,200 GB/s, and HBM4 targets 1,600–2,000 GB/s. Samsung announced mass production on February 12, 2026. If memory bandwidth grows about 1.5 times per generation while model parameter counts grow four to ten times per cycle, bandwidth pressure increases unless software or model design reduces memory traffic.
Supply concentration: Estimates put SK Hynix at 57–62% of the HBM market, Samsung at 33%, and Micron at 5%. Roughly 80% of supply originates in South Korea. Reported demand growth of 110–150% year over year, together with HBM4 price estimates of $500–550 per unit, makes price and supplier concentration part of the deployment constraint. Estimates that memory approaches 80% of GPU manufacturing cost should be treated as product- and configuration-specific rather than universal.
Physical constraints can therefore slow deployment without preventing it. The relevant forecast is how quickly each constraint changes, not whether it exists today.
The Citrini Argument: Concentration Creates Fragility
Citrini’s scenario starts on the demand side. It estimates that white-collar workers represent 50% of US employment but drive about 75% of discretionary consumer spending. It also cites the top 10% of earners as responsible for more than half of consumer spending and the top 20% for about 65%.
Under those assumptions, displacement among high earners has an outsized effect on services, real estate, and discretionary goods. Citrini illustrates the mechanism with a product manager whose income falls from $180,000 to $45,000 after moving to gig work. That example is part of the scenario, not an observed labor-market transition.
Citrini traces the transmission channels:
- Labor share of GDP: The historical series falls from 64% in 1974 to 56% in 2024. The scenario then assumes 46% by 2028, a 10-percentage-point decline in four years after an eight-point decline over the preceding 50.
- Real estate: The scenario puts year-over-year home-value declines of 8–11% in San Francisco, Seattle, and Austin, and places rising early-stage mortgage delinquencies in ZIP codes with more than 40% technology and finance employment. It applies those losses to a US residential mortgage market of about $13 trillion.
- Private credit: The report starts with market growth from less than $1 trillion in 2015 to more than $2.5 trillion in 2026. It then has Moody’s downgrade $18 billion of debt across 14 private-equity-backed software issuers.
- Fiscal: The scenario puts federal receipts 12% below the CBO baseline as payroll-tax and capital-gains revenue fall and unemployment benefits rise.
The scenario assumes median US token consumption of 400,000 tokens per person per day by March 2027, ten times its assumed level at the end of 2026. That premise requires agentic systems to move quickly from developer tools into commerce, financial planning, and real-estate transactions. The downstream losses depend heavily on that adoption assumption.
Citrini uses India’s IT-services sector as an international transmission channel. The sector exports more than $200 billion annually through companies including TCS, Infosys, and Wipro, with material exposure to software development, quality assurance, and business-process outsourcing. The scenario has the rupee fall 18% in four months.
Where the Arguments Intersect
The reports disagree primarily about timing.
Citadel argues that physical constraints slow adoption and that earlier productivity gains expanded consumption rather than collapsing it. It invokes Keynes’s prediction of a 15-hour workweek as an example of demand growing with productivity.
Citrini asks what happens if agentic AI moves rapidly from augmenting white-collar work to displacing it. Because its scenario concentrates spending and credit exposure among the affected workers and firms, the model produces correlated losses across asset classes.
The supply constraints in Citadel can coexist with the losses in Citrini. The result depends on whether deployment accelerates before labor, consumption, and credit markets adjust.
What the Constraints Model Shows
Working through the infrastructure math reveals several things both pieces understate:
Efficiency techniques materially change memory demand. Quantization, multi-latent attention, grouped-query attention, PagedAttention, and speculative decoding attack different parts of the inference path. Reported gains cannot simply be multiplied because they overlap and depend on the workload. In combination, however, production implementations can serve substantially more work than a naive FP16, multi-head-attention baseline. A forecast based only on raw hardware trends will miss that effect.
Price may limit deployment before physical availability does. In configurations where memory dominates accelerator cost, concentrated supply and rising prices establish a floor under inference cost. That floor affects which uses are economic and which organizations can deploy them at scale.
The median-token premise deserves scrutiny. Usage is highly skewed: a small group of power users can consume orders of magnitude more tokens than the median. Evidence about the tail does not establish the median. Citrini’s 400,000-token premise requires mass adoption, so the scenario should be tested against median usage and retained-user data rather than aggregate token growth.
The displacement-as-deflation argument has a missing term. Citrini models the demand destroyed by displacement but understates the supply response: cheaper intelligence can create new categories of demand. Electricity, computing, and the internet each did so. The unresolved issue is timing: correlated stress during the transition may trigger a financial crisis before new demand reaches equilibrium. Citrini’s argument therefore concerns transition dynamics rather than the steady state.
Infrastructure spending offsets some lost demand. Hyperscaler capital expenditure of $150–200 billion per quarter pays for construction labor, electrical equipment, land, and cooling systems. It can partly offset white-collar displacement, although the jobs arise in different places and require different skills. Because capital expenditure can fall, it is an offset rather than a permanent demand floor.
A Sequence to Test
The following order is a hypothesis to test, not a fixed hierarchy:
- Power delivery and grid interconnection: Reported lead times of 18–36 months depend on transmission capacity and utility planning. Software cannot remove this physical delay.
- Memory economics: HBM concentration and prices can make an otherwise feasible workload uneconomic. Quantization and architectural changes reduce memory demand but do not eliminate memory cost.
- Memory bandwidth: A workload can remain bandwidth-bound even when more FLOPS are available. The relevant measures are tokens per second, memory traffic per token, and utilization for a specified workload.
- Organizational adoption: Regulation, compliance, integration, process redesign, and user acceptance can delay deployment after the technology becomes economical.
- Labor-market adjustment: Citrini’s losses require the first four constraints to relax faster than workers, wages, consumption, and credit can adjust.
At the time of writing, power, memory cost, and bandwidth all restrict deployment. The risk case requires efficiency improvements and new HBM to relax the memory constraints while power capacity continues to grow. The empirical question is then whether deployment outpaces economic adjustment. Grid queues, HBM prices, workload-level utilization, enterprise deployment rates, and labor transitions would confirm or reject that sequence.
Bottom Line
Citadel presents evidence against a present crisis. Citrini shows how a crisis could propagate if its rapid-adoption and displacement assumptions hold. Memory bandwidth remains a technical constraint, while software efficiency keeps changing its severity. Power delivery may set the deployment schedule, and memory prices may decide which deployments are economic.
DRAM is therefore one constraint, not a complete explanation. Raw hardware specifications understate software gains, while supply concentration and pricing determine the cost floor those gains must overcome.
Watch HBM pricing and SK Hynix’s capacity announcements more closely than FLOPS benchmarks. The economics of memory, not the physics of compute, increasingly set the terms.