AI最大樽頸已經唔係GPU？真正卡住AI嘅，係「電」⚡

Core business message

• The “AI bottleneck” has shifted from compute hardware (GPUs, HBM, packaging) to physical infrastructure, especially:
  • Electricity availability
  • The operational timeline to make power usable for AI workloads.
• The key KPI is Time-to-Power: the time from deciding on a data center project → actually getting power → turning that into computing capacity → ultimately converting into free cash flow.
• Investors and operators should stop treating “X GW capacity” as equivalent to “X GW delivered,” and instead evaluate:
  • Delivery certainty (grid interconnects, permits, long-lead equipment)
  • Execution speed

---

KPI / framework introduced: “Time-to-Power”

Definition

• Time-to-Power = time from project decision → power delivery/spend readiness → computing capacity utilization → cash flow generation.

Business logic / why it matters

• Delays are financial (NPV impact), not just engineering schedule slips.
• If cash flows start later, the discounted value drops materially.

NPV example given

Two identical data centers:

• Start at time A: NPV = $7.6B
• Cash flows pushed back 2 years: NPV = $6.3B
• NPV loss ≈ 17%

Additional delay effects mentioned:

• Labor/contractor cost inflation
• GPU/equipment cost changes by delivery year → The business case changes because Time-to-Power itself changes the economics.

---

Operational playbook / “project pipeline” model

The speaker frames an AI/data-center project as a staged conversion funnel:

1. Project decision
2. Power availability
3. Equipment procurement (GPUs, etc.)
4. Construction
5. Full power supply
6. Free cash flow

Rule:

• Any delay anywhere shifts the entire timeline backward, harming financial outcomes.

---

Physical bottlenecks and what to verify (execution checklist)

The video repeatedly contrasts “capacity claims” vs deliverability. Key items to validate:

• Grid realities
  • Grid connections already secured?
  • Interconnect status with the grid operator?
  • Substations/transformers capacity available?
  • Transmission upgrades required—and can they be delivered fast enough?
• Permitting + approvals
  • Permits in place?
  • Approvals completed vs still pending?
• Construction + FID
  • Which stage is the project at?
  • Is it past FID (Final Investment Decision)?
  • Timeline to reach phase milestones needed for FID?
• Long-lead equipment
  • Switchgear / transformers / gas turbines availability
• Power deliverability vs generation
  • Generation ≠ deliverable capacity if the grid can’t accept it

---

Market/infrastructure demand context (high level)

• Projected datacenter electricity consumption:
  • ~485 TWh (2025) → nearly ~950 TWh (2030)
• AI datacenters may be even higher (especially in the US).
• By 2030, datacenters could represent ~one-third of increased US power demand.
• The constraint is physical: transmission/substations/transformers/permitting/construction (software changes don’t solve this).

---

Concrete case studies (execution strategies)

1) Microsoft + Constellation Energy: grid deliverability vs generation

• Constellation preparing to restart a reactor at Three Mile Island Clean Energy Center.
• Microsoft supports via a PPA (power purchase agreement).
• Target: reactor could be back online by March 2028.
• Key point: Time-to-Power is constrained by grid operator PJM.
  • Power may exist, but the grid might not be ready to deliver it.
• Execution insight:
  • Partial capacity interconnect / earlier licensing at the site can reduce the total interconnection burden (not everything needs to be handled simultaneously).

2) Google Finland: nuclear supply + PPA visibility

• Google plans €1B investment in data-center-related infrastructure over two years.
• Mentions a major nuclear-linked project designated 13 GW HLPGF.
• After a 2022 PPA, start targeted for 2028.
• Aim: 50% capacity increase of the 2030 power plant (referenced as “49V” in subtitles; likely a specific facility name).
• Business takeaway:
  • PPA provides time visibility and can reduce some execution uncertainty.
  • But this is part of a broader investment program—risk analysis should be staged (speaker references splitting investment decision risk across phases).

3) Microsoft (alternative risk reduction): build/own generation integration

• Microsoft signs a 20-year power agreement in Texas.
• Reaches about 2.67 GW.
• Stated targets:
  • Goal: 2028
  • FID targeted for late 2026
• Strategy shift:
  • Previously: datacenters requested power from the grid.
  • Now: power generation becomes part of the data-center strategy, reducing execution risk tied to external grid scheduling.

4) Oracle (financing/capex risk transparency lens)

• Q1 2027 performance figures mentioned:
  • Application revenue: $64B
  • AI cloud contracts: >$30B
  • Operating cash flow: ~$23.1B to $28.5B
• Payment mechanics:
  • Oracle received $11.36B in customer payments.
  • Because these payments are already part of operating cash flow, the speaker treats them as an additional indicator connected to Gross CapEx.
• CapEx risk framing:
  • Net CapEx cited as about ~$1B (after accounting for customer payments/related costs).
• Key investment conclusion:
  • Don’t only look at “how big capex is”—evaluate who funds it and whether the company de-risks capex via customer payments or project financing.
  • Analogy: judge AI “block quality,” not just block count.

---

Metrics / KPIs explicitly cited

• Time-to-Power (primary KPI)
• NPV impact
  • $7.6B → $6.3B with 2-year delay (~17% NPV loss)
• Electricity demand
  • ~485 TWh (2025) → ~950 TWh (2030)
  • Datacenters could be ~one-third of US power demand growth by 2030
• Oracle
  • App revenue $64B
  • AI cloud contracts >$30B
  • Operating cash flow $23.1B–$28.5B
  • Customer payments $11.36B
  • Net CapEx ~$1B (per speaker’s framing)
• Commitments / liabilities (proxy for capacity locking)
  • Future payment obligations: $1.09T (since August; scale of commitments)
  • Oracle lease commitments: $288B, mostly for data centers operational 2027–2029

---

Actionable recommendations for investors / operators

• Replace “capacity owned” with “power delivery timeline”
  • Ask: When will power actually be available?
  • Check: permits, FID, grid connection, transformers, long-lead equipment.
• Measure project certainty + financing quality
  • Evaluate whether capex risk is borne by the company vs shifted through customer payments / financing.
• Use staged underwriting
  • Don’t underwrite as if a “GW headline” will translate to delivered capacity on day one.
  • Demand evidence of milestones (interconnection, licenses/amendments, construction readiness).
• Watch for fixed-cost risk from locked commitments
  • Capacity may be “locked,” but if power delays/construction setbacks/missed demand targets occur, fixed liabilities can become costly.

---

Framework explicitly stated: “Infrastructure quality”

Speaker’s formula:

Infrastructure quality = demand quality + project certainty + funding quality + speed of conversion

---

Presenters / sources mentioned

• Microsoft
• Constellation Energy
• PJM (grid operator referenced)
• Google
• Oracle
• Meta and Alphabet (mentioned in the commitments context)
• Investor perspective (no single named investor presenter)
• Three Mile Island Clean Energy Center (project site referenced)