Daily Alpha · Substack
· Post-Market Alpha · by Buzzberg Research
The neocloud thesis is shifting from GPU ownership to power-secured capacity and tenant prepayments, while token demand runs far ahead of installed compute.
Themes on this desk
Token demand vs compute
Outlier Capital cites OpenRouter token usage near 380T/month with agents ~70% of tokens, against installed compute rising 4x by 2028 versus a 24x token-demand implication.
Neocloud economics
IREN, Cipher, Core Scientific, WhiteFiber, HIVE, CoreWeave and Nebius are graded by contracted power, tenant prepayments and leverage.
OpenRouter token usage inflected to ~380T/month by August 2026
Outlier Capital cites a Citi chart showing OpenRouter token usage went from almost nothing in January 2025 to close to 380 trillion tokens in August 2026, a 31% month-on-month CAGR and +2,434% year-over-year.
If token consumption is compounding at ~31% monthly, the binding constraint shifts from model capability to physical compute and power, favoring owners of contracted capacity over application-layer names.
Watch Subsequent monthly OpenRouter or comparable token-volume disclosures; a break in the month-on-month growth rate would undercut the shortage thesis.
Source →Compute grows 4x by 2028 while token demand implies 24x
Citing Morgan Stanley, the author notes hyperscaler data-center spend of $1.5 trillion in 2027 (+60%) slowing toward ~12% growth in 2028, with installed compute rising from 35 GW in 2025 to 145 GW in 2028 — 4x compute against Goldman's 24x token demand.
The arithmetic residual is the author's stated shortage; if correct, pricing power accrues to whoever controls interconnect and energized megawatts, not to whoever owns the most GPUs.
Watch Actual 2027-2028 hyperscaler capex prints and installed-GW disclosures versus the $1.5T and 145 GW figures.
Source →IREN: full-stack ownership with Microsoft $9.7B/200MW contract
The author argues IREN is the only listed name owning land, interconnect, substations, liquid-cooled halls, GPUs and software, with Microsoft as a $9.7B/200MW five-year contract, Nvidia on the same campus, $4B contracted ARR against 2026 capacity and $1B operating ARR after Horizon 1; he assigns ~50% 5x probability and says he initiated at $40.8.
If IREN is valued at roughly one-third the equity value of CRWV/NBIS with comparable bottleneck exposure, the re-rating from miner to AI-cloud multiple is the core asymmetry the author is underwriting.
Watch Horizons 2-4 delivery by December, Microsoft prepayment terms, and whether Bitcoin mining actually shuts down by year-end as stated.
Source →Tenant prepayments change the funding math for GPU capex
The author states Microsoft prepaid ~20% of IREN's contract and that recent deals prepay 45-55% of GPU capex, funded by the tenant and asset-backed paper tied to the tenant.
Prepayment shifts the capital burden off the operator's balance sheet, which is the key differentiator versus lessors funding racks through equity issuance or ATM programs.
Watch Disclosed prepayment percentages in future NeoCloud contracts; a shift toward operator-funded capex would compress the advantage.
Source →CoreWeave: ~$104B backlog but leverage caps the 5x case
The author calls CoreWeave the finished category winner with ~$104B backlog from Microsoft, OpenAI, Meta and Anthropic, but flags debt-to-equity in the hundreds of percent, enterprise value roughly double equity, and assigns only ~15% probability to a 5x in five years.
A leveraged GPU lessor priced as if the identity change is complete offers less asymmetry than the author's preferred names, framing CRWV as a quality-but-crowded holding.
Watch Backlog conversion into billed revenue and any change in financing terms or interest burden; a 5x would require roughly a $225B equity value.
Source →