Video summary
[데이터센터 시리즈] “칩 부족 끝나면 이게 부족해집니다” AI 데이터센터 2030년까지 돈 몰릴 ‘다음 병목’
Main summary
Key takeaways
Core thesis: AI data centers “next bottleneck” after chips
- The “chip shortage” won’t be the final constraint on AI infrastructure.
- After semiconductors, the next binding bottlenecks become:
- Power generation & grid interconnection
- Licensing/regulatory approval
- Fiber/subsea connectivity (increasingly critical as sites move farther from existing network hubs)
- In the U.S., even when companies can afford chips and equipment, licensing delays can block execution—widening the gap between firms that already have approved sites + power sources versus those without.
U.S. site selection playbook: 5 conditions
- Power
- Cooling
- Site (land/region suitability)
- Fiber optic cable connectivity
- Licensing (described as an emerging key bottleneck, especially when approvals are time-compressed)
Stakeholder/market reality check: community opposition
Survey-based claims include:
- A poll of 7,000 adults (Aug 20–Sep 1):
- 69% oppose constructing an AI data center nearby
- 45% strongly oppose
- The Economist survey (late August):
- 47% of Trump 2024 voters oppose
- 35% support
- Policy backlash trend:
- Mentions increasing state-level opposition, including New York (July) and Texas (August).
Example: Texas “power request” saturation
- Texas reportedly reached 474GW of power requests—more than 5× the all-time peak.
- The implied takeaway: speed and scale in requests don’t translate into executable capacity fast enough.
Why firms may move outside the U.S. (execution strategy shift)
- If U.S. licensing takes too long, companies pursue:
- Regions with faster approval dynamics, and/or
- More favorable natural/economic conditions
- Underlying premise: AI infrastructure expansion must continue, framed as monetizable and strategically necessary.
National-defense framing (strategic operations implication)
A cited statement (Sep 8) argues that if China wins in AI, there’s “no later” (context includes defense budget >$550M and an “Iron system”).
- Data centers are positioned as fundamental AI industrial infrastructure
- The concept expands from commercial operations into national security capability
International expansion examples (what each region “solves”)
Patagonia, Argentina (power + cooling + environmental setup)
- Cited backdrop: Reuters coverage of plans for 500MW data centers
- Additional regional/company mentions:
- Neocan (Argentina) with 120MW
- 300MW in Poland mentioned alongside broader site pipelines
- OpenAI’s $25B announced project (Oct, prior year): described as early-stage with no confirmed construction start
- Claimed advantages:
- Low temperatures due to high altitude → lower cooling costs
- Steady winds (wind potential)
- Hydropower
- Empty land + favorable environmental conditions
- Implied caution:
- Even if a region is attractive, starts can lag (delays, governance, and investment-to-execution translation issues).
Australia (demand pull + investment pipeline)
- Bloomberg-reported sentiment (Sep 7):
- Operator AirTrank (Sydney HQ) cites stronger-than-expected messages from Google, Apple, Meta, Amazon, Microsoft over 6–8 months
- Reasons cited:
- Political stability
- Renewable energy access
- A “usable party” phrase (unclear; likely meaning regulatory/business readiness)
- Investment figures:
- Microsoft CEO Nadella reportedly visited Sydney
- ~$18B investment by 2029 (Apr 23)
- Macro impact claim:
- Commons Bank estimate: by 2030, data-center construction contributes 6% and 5 percentage points to real corporate investment growth rates in 2026 and 2027 (respectively)
New Zealand (grid carbon profile + stable renewable + water-safe cooling design)
- Electricity mix:
- >90% low-carbon
- ~1% geothermal (described as stable 24/7 vs weather-dependent renewables)
- Example project:
- ~250MW data center (IT load) planned at/near a closed gas power plant
- Closed-loop cooling to protect local water resources
- Battery integration for resilience/backup
Communication bottleneck: fiber + submarine cables + satellite “backup”
Fiber rollout time as a secondary bottleneck
- Even if computation is available, sites must connect to communication networks.
- Submarine cable demand rises as sites move into more remote/long-distance geographies.
Concrete cable market facts
- “Currently” on Earth: 500+ commercial submarine cables
- 99%+ of intercontinental internet traffic via submarine cables
- Hyperscaler bandwidth share:
- Historically small (as of 2010)
- Now >70%
- New submarine cable investment (2026–2029):
- Total planned value > $16B
- Example ownership/involvement:
- Google has stakes in 30+ systems
- Specific example: Humboldt Cable
- 14,800 km
- Route: Chile → French territories → Sydney, Australia
- First cable directly connecting South America to Asia-Pacific
- Operations begin 2028
- Cost structure claim: Google covers 99% of costs; local infrastructure covers 1%
Risk + backup channel
- Submarine cables face disruption and concentration risks (e.g., damage and “Red Sea system” reference).
- Satellites are framed as next-gen communications, but not full replacements due to bandwidth/speed limitations.
- Examples:
- Starlink (SpaceX)
- Amazon collaborating with AT&T on satellite communications
- Implication: hyperscalers pursue parallel communication modalities—fiber now, satellites for resilience and future capacity.
Actionable “capex stack” direction to 2030
1) Power generation facilities (onsite generation; don’t rely only on the grid)
- Meta “Albata” example
- 1GW data center + 932MW gas power plant
- Long-term plan: expand regional supply
- Korean supply chain examples:
- HD Hyundai Electric:
- Long-term supply contract (July 2) for >1 (subtitle truncated; implied to be “over $1B” or over 1GW-scale data center power distribution for North America)
- Sequential supply starting 2028
- Hyosung Heavy Industries:
- Long-term contract for ultra-high-voltage transformer reactors with Osnet (Victoria, Australia)
- Order value: exceeds 3 billion won
- Supply: 5 years
- Reason: Australia grid overhaul → transformers needed regardless of specific data center location
- HD Hyundai Electric:
2) Connectivity infrastructure (HVDC submarine cables + cable manufacturing capacity)
- LS Cable & System reportedly quadrupled HVDC submarine cable production capacity.
- Note: power submarine cables and communication submarine cables are different products (manufacturers may overlap, but are not identical).
- If data centers diversify globally, demand increases for both power and comm cables.
3) Space/orbit (long lead-time but strategic optionality)
- “Thorin AI” is mentioned as a framing of tech direction (unclear exact term).
- Starlink/Starship commercialization assumptions:
- Costs need to drop
- Starship commercialization is treated as a prerequisite
- Satellite/space tech is framed as relevant through ~2030, positioned as more than “just investment ideas.”
4) Energy-sector investing signals (high level)
- Mentions Peter Thiel’s 13F: large allocation to electricity/power/energy
- Includes references to:
- Investment in an Argentine energy company
- Other energy-related companies (including utilities/energy firms and an “X Energy” small reactor company)
Metrics / KPIs explicitly mentioned (and how they’re used)
MW scale examples
- Patagonia: 500MW
- Neocan: 120MW
- Argentina pipeline: another 500MW scale mention (with 300MW in Poland referenced as context)
- New Zealand: 250MW data center (IT load)
- Meta Albata: 1GW data center + 932MW gas plant
Power oversubscription / queueing
- Texas power requests: 474GW
- U.S. licensing described as a slowing factor (no single KPI provided, but “gap will widen” is the conclusion)
Investment totals / timelines
- OpenAI project: $25B announced (early stage; no construction start yet)
- Microsoft Australia: ~$18B by 2029
- Subsea cable plans (2026–2029): >$16B
- Humboldt Cable:
- Operations begin 2028
- 14,800 km
Survey opposition (community pushback)
- 69% oppose, 45% strongly oppose (Aug 20–Sep 1 poll; n=7,000)
- The Economist survey: 47% oppose / 35% support among Trump 2024 voters
Communications share
- Hyperscalers: intercontinental bandwidth >70% (as claimed)
Macro investment contribution (Australia)
- Data centers’ contribution to real corporate investment growth:
- 6% and 5 percentage points to growth rates in 2026 and 2027 (respectively), per the cited estimate
Risks and execution pitfalls (what delays growth even with capex queued)
- Delays likely outside the U.S. too
- Even if companies move, other countries may face similar opposition and political risks.
- Argentina is used as an example, including political risk and execution lag.
- Bottleneck type may be physical rather than “chip availability”
- Semiconductor constraints can ease, but shortages of land and power facilities can’t be “pushed out” as quickly as chip supply.
- Orders vs revenue growth
- Orders may accumulate while revenue growth stays slow due to permitting, power delivery timelines, and grid/cable build cycles.
- Commodity indicator (weakly framed)
- Copper prices at all-time highs used as a signal for infrastructure buildout intensity and related bottlenecks.
- Near-term investing caution (high level):
- Forecasting is difficult; emphasis on composure/risk management (framed more as general investing commentary than business execution detail).
Key frameworks / playbooks referenced
- 5-factor site selection framework
- Power, Cooling, Site, Fiber connectivity, Licensing
- “Bottleneck stacking” logic
- Chips → power → licensing → fiber/subsea → satellites as contingency
- Scenario planning until ~2030
- Treat comm/energy/space buildouts as parallel systems required for sustained AI infrastructure growth
Presenters / sources mentioned
- Reuters (Patagonia Argentina / 500MW coverage; Sept 7)
- Bloomberg (Australia sentiment; Sept 7)
- The Economist (late August survey)
- MBC News (referenced for Patagonia context)
- Finance Minister Bessent (quote; Sept 8)
Companies/entities mentioned
- OpenAI
- Patagonia / Neocan
- AirTrank
- Microsoft
- Amazon
- Meta
- Apple
- Implied by messaging list: Oracle
- Starlink (SpaceX)
- AT&T
- LS Cable & System
- HD Hyundai Electric
- Hyosung Heavy Industries
- Osnet
- Commons Bank
- Thiel / 13F
- Elon Musk
- Plain Labs
- Nvidia (earnings cited for “demand shown”)
- X Energy
- Humboldt Cable (project example)