SHP AI

Building the utilities of tomorrow

With the fastest time to compute and capital efficiency

October 2026

2026: Parabolic AI demand meets power wall

OpenRouter tokens up 19x in 8 months

Tokens per month (trillions)

80 GW US power deficit by 2030

US data center power gap (GW)

Ecosystem is supply constrained, with more chips than powered real estate

Demand: OpenRouter tokens per month. 2025 points as in v2 (21.5, 28, 35 trillion), except January 2025 at 2 trillion (Alan, 28 Oct: about 0.5 trillion a week, times 4). 2026 refreshed from OpenRouter's rankings page (weekly token totals, all models; downloaded 28 October 2026): average of the full weeks starting in each month, times 4.3. January 32 trillion; February 54; March 87; April 94; May 122; June 190; July 248; August 401; September 610. Up 19x from January to September. Anthropic's annualised revenue went from US$9bn in December 2025 to US$65bn in July 2026. Anthropic's S-1 (26 October 2026): Q3 2026 revenue was 86 times Q4 2024.

Talking point (Alan, 28 Oct): introduce Anthropic verbally. Most of its revenue, about 80%, is pay per use (intelligence as consumption), and subscriptions are about 20%, so tokens show demand most directly.

Talking point: US low hanging fruit is gone. Power: BCG, a 10 GW surplus in 2024 becomes a gap of up to 80 GW by 2030. Primary US market vacancy is a record low 1.4% (CBRE H1 2026); grid connections take more than five years (LBNL, July 2026).

Chips versus shells: on the BG2 podcast with Jensen Huang (15 October 2026), Altimeter's Clark Tang put 2027 at about 50 GW of announced chips against 35 to 40 GW of powered land and shells coming online. Jensen Huang: "Land Power Shell is clearly one of the obstacles."

Emerging markets as a source of compute

Existing power

EM has material volume of existing power assets

4,772

Special economic zones
in developing countries

Energization speed

Advantageous in many jurisdictions,
with popular opinion supportive

Majority

Supportive vs. 75% in US
oppose building data centers

Suitable workloads

Regional and global async workloads

~50%

of frontier lab compute
used for post-training

Increasingly attractive paradigm with no more low lying fruit in the US

Flow: the previous slide is the problem, parabolic AI demand meets the power wall, so compute is supply constrained. This slide is the answer: turn local electricity in emerging markets into compute sold globally. Three reasons it works.

Power. There is a lot of power in emerging markets that is available now and can be converted: the low-hanging fruit. UNCTAD counts 5,383 special economic zones worldwide in 2019: 4,772 in developing economies, 374 in developed economies (262 in North America) and 237 in transition economies (World Investment Report 2019, Table IV.2). Parks already have the plant, the substation and the land.

Workloads. Demand sits in the US, but much of it does not need to run there. Post training and RL are now about half of frontier lab compute: SemiAnalysis (Tokenomics Model, "Long Live the Short King: Why 4-hi HBM Wins", 13 Sep 2026) puts Anthropic and OpenAI combined at 50% post training / RL, 37% inference and 13% pre-training in 2Q26, and 55% / 38% / 7% in 4Q26E. A year earlier the mix was the other way round: in 2Q25 post training was 14% and pre-training 50% (same SemiAnalysis chart), so the two have basically inverted. These jobs are asynchronous, so latency to the end user does not matter (Prime Intellect, 15 Apr 2025: RL is "inherently more asynchronous" than pre-training). Anthropic's S-1 (26 Oct 2026) directs capacity "regardless of provider, chip architecture, or physical location" and is "pursuing data center capacity in additional jurisdictions outside the United States".

Speed to energization. Under one year from origination to compute on an existing powered site, against a US median of more than 5 years from grid request to commercial operation (Lawrence Berkeley National Laboratory, Queued Up, 1 Jul 2026). Public opinion is a tailwind in emerging markets and a headwind in the US. Public First, AI Global 2026 (for ITIF, Aug 2026), "Should your country build more or fewer data centers?": build more Nigeria 74%, Vietnam 64%, India 65%, Indonesia 56%, Brazil 48%, South Africa 48%, Mexico 40%, a simple average of 56% (9% build fewer); US 26% more, 24% fewer. Heatmap Pro poll by Embold Research (2,045 voters, 8-13 Aug 2026): 75% of Americans would oppose a new data center near where they live; more than 70% still did in October. Note: the EM figure asks about building more nationally, the US 75% about a data center nearby.

Evidence file: deck-v3/research/s-local-to-global-evidence.md.

We originate the sites that power our compute

1Sources readily available power assets to monetize

Originates and funds de‑risking until power is secured
Taken out and earns residual and compute upside

2Purchases site once power secured

100% owned by compliant capital
Funds remaining equity development capital

3We lease and operate, and arrange rest of the stack

Sign the Compute Buyer, for whom the site is purpose built
Arrange financing (construction and GPU)

Origination and Equity Partners systematically produce new AI colocations for us

Our Origination Partners are Chinese industrial groups that already hold power assets abroad. They put their own capital at risk to take a site to power secured, and they are paid for de-risking it.

They earn about 4x on that capital within two years and about 10x within five, through the development return at financial close, their share of the residual and their share of compute profit. For them it is a way to monetize power assets they already hold.

Once power is secured, a Western Equity Partner takes out 100% of the de-risked site at financial close and funds its development. Precedent: Blue Owl took 80% of Meta’s Hyperion campus at financial close (Meta, 21 Oct 2025).

SHP AI brings the parties together, takes the colocation lease and shares the compute profit upside. Compute earns about 8-10x the shell rent per MW: about US$20-25mm of compute revenue a year (Nebius) against about US$2.4mm of shell rent (Bitdeer). We repeat this site after site.

Power to compute: The fastest builds, with capital efficient GPUs

4-6 month builds, vs 8-9 months at Fluidstack
Buyers and lenders fund 92% of our GPUs, or GPU Partners own them
Take or pay wholesale contracts with labs and hyperscalers
Site and GPU risk sit with the investors who want it

Fastest power to compute and capital efficient GPUs enable rapid scaling

1. Sites are purpose built and leased wholesale to large scale compute buyers: frontier labs and hyperscalers. They buy end compute from us.

2. Fastest power to compute, leveraging state of the art engineering for 4-6 month builds.

3. The GPUs are owned either by us or by a GPU Partner; the two are interchangeable. When we own them, buyer prepayments and GPU debt fund most of the chips. When a GPU Partner owns them, institutional partners structure, finance and deploy them off our balance sheet. GPU Partners are separate from the site Equity Partners, because their risk appetites differ.

4. We are aligned with the GPU equity holders: they share in the compute profit upside.

5. Each source of capital gets the risk and return it wants: Equity Partners own the powered shell, GPU Partners own the chips, and we run the compute.

Proof points. Build: 4-6 months against 8-9 months for Fluidstack at Lake Mariner CB-4, first steel to first hall, in halls of 100-170 MW (Epoch AI, 15 July 2026, via v2 research). GPUs: on the first site, US$2,040mm of GPUs; buyer prepayments (US$743mm) and GPU debt (US$1,224mm) fund 92% of the first tranche, and our peak equity is about US$55mm (model v9.23). With a GPU Partner the chips sit off our balance sheet. Contracts: take or pay, 10% of capacity on one-year, 35% on three-year and 55% on six-year terms (model).

Historic premia for compute

Monetization

(US$m per MW IT a year)

GPUs pay back in 1.5-2 years

US$23-28m per MW a year from a short-term and 6-year mix, against ~US$30m of GPUs

Every partner shares the premium

20% of compute profits shared with Equity Partner and Origination Partner
Locks in alignment for speed, scale and long term success

New leases earn 2x the old book

Incumbent neoclouds earn ~US$9m on old contracts; new 3-6 year leases pay US$15-20m

Rich premiums enable GPU financing

Off balance sheet financing splits cash flows; without rich yields there is nothing to split

Today's scarcity premium funds rapid scaling

Cost side: about US$30m per MW IT all in if we rent the shell and buy the chips, and US$5m to US$7m a year of cash opex. Monetization side, annualised: SpaceX took US$30m to US$50m per MW a year from Anthropic on short-term capacity; the neoclouds, CoreWeave and Nebius, average about US$9m on their historical books; the colocation providers, Digital Realty, Equinix and CoreSite, make about US$4m. The dotted lines are today's benchmarks: self-serve at about US$100m, short-term bare metal at US$40m to US$50m, three-year leases above US$20m, six-year Anthropic-style leases at US$14m to US$15m.

Payback (Alan, 28 Oct): blending short-term and six-year contracts monetizes at US$23-28m per MW a year, so about US$30m of GPUs per MW pays back in 1.5-2 years. Long contracts make the asset financeable; short contracts pay it back.

The premium is shared: 20% of compute profits is shared with the Equity Partner and the Origination Partner, so every party is paid to move fast. Earlier sellers locked in low prices on long contracts; we sell into today's scarcity, and prices will normalize later.

Rich premiums make off balance sheet GPU financing work: it splits the cash flows between the GPU Partner and us, and without rich yields there is nothing to split.

Enabling win-wins across the ecosystem

Compute BuyerUS$130mm / MW of revenue on US$30-50mm of cost

Origination PartnerFast capital cycle and a share of the compute premium

Equity PartnerRich returns on de-risked, scarce assets

Private creditRicher yields, growing AUM

SHP AIScale fast into capital efficient token foundries

400 MW of sites already under exclusive lease or MOU, sourced by founders of 5 listed companies

Pipeline (MW grid)

400Exclusive lease or MOU, 5 sites

100Power secured by end 2026

200+Ready for service in 2027

Origination Partners

5Listed company founders

3Families with controlling stakes

Power: 2 TWh sold a year

Cooling: 30+ data centers delivered

Servers: plant for 100,000 servers a year

On track to deliver 200+ MW of compute and 1 GW of new origination in 2027

MW Pipeline: Sites 1 to 5 add up to 401.5 MW grid. Site 1 is 37.5 MW grid in two phases (12.5 + 25) with its grid approval; Sites 2, 3 and 4 are about 100 MW grid each; Site 5 is 64 MW grid in two phases, with power permits held by the owner. Locations are confidential.

2027: on the current ETAs, 219.5 MW grid reaches ready for service by end 2027: Site 1 both phases, Site 2 phase 1, Site 3 and Site 5 phase 1. Site 4 follows in Q1 2028.

Origination Partners: the chairmen and founders of five listed companies, plus three families who own major assets or are controlling shareholders. They source the sites, fund pre-development and bring power and build expertise. Names under NDA only.

Power secured: Site 1 grid approvals (Oct 2026) and the Site 5 owner's permits, 101.5 MW grid by end 2026 on the site tracker. Partner proof points are from the v2.11 traction slide: the power partner sells 2 TWh a year to about 1,000 industrial customers; the cooling partner has delivered 30+ data center projects; the server partner is building a plant for 100,000 servers a year (early construction).

1 GW of new origination in 2027 is on top of today's pipeline; it accelerates after the seed closes.

Landscape: Token foundry vs. neocloud

Token foundry

Neocloud

Capital

Off balance sheet

100%+ loan to cost on chips

On balance sheet

Equity and GPU-backed debt

Price

US$m per MW IT
a year

14-20

3-6 year contracts today

~9

Locked in at historical prices

Speed

Years

<1

Site to compute

2-3+

US grid connection alone

The time is now for a new paradigm

Three differences between us, a token foundry, and a neocloud: capital, price and speed. Behind the capital row: financial markets have matured, AI compute went from corporate capex to an institutional asset class in 3 years, and 100%+ loan to cost and off balance sheet financing are now possible in 2026.

One: we turn power and capital into tokens efficiently; a neocloud raises equity to sell compute. CoreWeave priced its IPO on 27 March 2025, 37.5 million shares at US$40, about US$1.5bn.

Two: we capture today's compute premia; neocloud books are locked into older contracts. CoreWeave and Nebius average about US$9m per MW IT a year on their historical books, against US$14m to US$20m for 3-6 year contracts signed today. Short-term capacity, US$30m to US$50m, is upside on top.

Three: financial markets matured. CoreWeave's first GPU facility in August 2023 was US$2.3bn of private credit; in June 2026 Apollo and Blackstone launched a US$35bn tranche for Anthropic's compute.

Four: 100%+ loan to cost and off balance sheet are now possible. CoreWeave's DDTL 4.0, March 2026: 90% LTC in installation, then 1.2x DSCR after customer acceptance, which CoreWeave says implies up to about 102% LTC.

Five: we take a site to compute in under a year. In the US, BCG says grid connection queues alone "add a minimum of two to three years"; Lawrence Berkeley Lab puts request to commercial operation at more than 5 years. Then come stick built halls and local opposition. November 2025, CoreWeave cut its 2025 revenue guidance to US$5.05bn to US$5.15bn, citing "a third-party data-center developer who is behind schedule".

Team: Founder

Alan X. Chen

Canadian citizen, native in English and Mandarin

West: Oxford, London and Canada

JPMorgan Structuring: Structured and illiquid credit, bank capital, rates and FX derivatives and principal investing. Helped build Scandinavian franchise from scratch, and was tasked with reinventing multi-trillion USD CDO market.
Special situations PE at $100Bn USD SWF manager, ran deal for acquisition of $1Bn loan book with securitization arbitrage that increased IRR by 1000bps.
Oxford Financial Economics (Barclays Capital David Newton Memorial Prize), 2012 Olympic torch bearer and 1 of 8 selected to celebrate with the Queen.

East: 11 years building in Beijing

Valeon Group (Founder): Achieved 500mm RMB+ in revenues and global #1 in helping Chinese students obtain Ivy League admissions (2,000+), powered by 2,000+ mentors including 50+ Rhodes Scholars, 250+ full time staff across four offices. Raised ~20mm USD, including from former global Chairman of McKinsey/Rolls-Royce, Head of Goldman Sachs Asset Management, founder of ABRY capital.

Team: Co-founders

Wendy Wang 王晓雯

Built, leased and sold a 3,300P H100 and B200 GPU cluster: Live within a year, 99.95% uptime against a 99.5% SLA, bought outright by a listed compute company after six months.
Delivered a 100 MW green direct-connect power system and a 3 GW wind portfolio; led the full-cycle investment and build of two large compute clusters.
Fixed income at two leading Chinese brokers: Bond underwriting, debt and structured finance.

Terrence Lan 兰天

Guosen Securities investment banking for 10+ years: Led RMB 20bn+ of bond, MTN and private placement financings, listed-company M&A and IPOs.
Founder of payments and compliance fintech Loyal, with RMB 5bn+ of transaction volume; clients include Ping An, UnionPay, Alibaba Cloud, BMW, Audi, Volvo, etc.
Built, leased and sold a 3,300P GPU cluster; investor in data centers, liquid cooling and AI systems integration.

Roadmap to 2027 year end

When	Areas of focus	Milestones
Q4 2026	Seed and setup	Seed round first close Corporate and admin set up Bring on board 2+ door opening advisors 0.5 GW of sites in exclusivity for development
Q1 2027	Hiring and pushing power to compute	Initial hiring for capabilities and relationships (clients, cloud, product, financings) across New York, Bay Area and Singapore Kick off engagements with tier-1 clients and financing partners Primary Equity Partner for takeouts confirmed
Q2 2027	Financial close and pushing for delivery	Complete first financial close Fundraising for Series A Power secured on 0.3+ GW in sites
Q3 2027	Series A and senior bench push	Close Series A Executive bench round out 0.2+ GW at financial close
Q4 2027	Compute at scale and revenue ramping up	US$0.5bn+ ARR from 0.15+ GW of sites ramping up 0.4+ GW of sites with power secured

Financials: Rapid scaling to blue chip-like EBITDA

Base case	2027	2028	2030	2032	2034
IT capacity billing, year end (GW)	0.00	0.06	0.82	1.57	2.33
Compute revenue (US$mm)	0	211	10,839	22,894	33,206
Compute EBITDA (US$mm)	0	152	7,700	15,926	22,478
Combined net profit (US$mm)	(4)	141	2,290	4,702	6,598

Base: 3 GW grid in 5 years, then 600 MW a year

Prices: New contracts fall to US$15mm per MW IT a year by 2034

Compute revenue reaches US$33bn in 2034 on 2.3 GW of billing capacity

Fundraise and use of proceeds

60%Team and hiring

20%Corporate structuring, export control,
tax, legal and compliance

20%Travel and operations

The seed takes us to Site 1 financial close

Why now: 2026 is the critical year

Hyperscalers and labs look abroad for power

80 GW US gap by 2030

Scarcity pricing pays for the build

1.5-2 year GPU payback

Capital now funds chips off balance sheet

US$35bn chip deal in 2026

Spare slide, held in reserve. Three reasons 2026 is the year; the detail is what I say, the slide shows one proof for each.

One, hyperscalers and frontier labs are starting to look abroad. Low hanging fruit at home is gone. Political opposition became a headwind in 2026: 75% of Americans oppose data center construction (Heatmap Pro / Embold Research, August 2026). And the US power gap reaches up to 80 GW by 2030, from a 10 GW surplus in 2024 and a 15 GW gap in 2026 (BCG, Solving the US Data Center Power Crunch, March 2026).

Two, compute premia are at historic highs. Demand went parabolic in 2026, powered shells are the bottleneck, and there are more chips than space to put them. Today's rates per MW IT a year, by contract term: about US$14-15m on six-year leases, above US$20m on three-year leases and US$40-50m for short-term bare metal. The chips cost about US$30m per MW IT all in. Blending short-term and six-year contracts brings in about US$23-28m per MW a year, so the GPUs pay back in 1.5-2 years; all on six-year leases at about US$16m, it takes about five years.

Three, compute financing markets matured in 2026: 3 years from capex to off balance sheet financing. CoreWeave's first GPU-backed loan was US$2.3bn in August 2023. In June 2026 Apollo and Blackstone funded a US$35bn vehicle that owns Broadcom chips leased to Anthropic, off the user's balance sheet. In August 2026 NVIDIA's platforms with Apollo, KKR and others targeted US$500bn of third-party capital.