Samsung Bets $1 Billion on AI Infrastructure: Why Power Is the New Bottleneck

On September 29, 2026, Samsung Electronics and five affiliates announced a combined USD 1 billion investment in Helix Digital Infrastructure, an AI infrastructure company established by KKR. Samsung Electronics contributes USD 500 million, with Samsung C&T, Samsung SDS, Samsung SDI, Samsung Life Insurance and Samsung Fire & Marine Insurance funding the remainder. Helix was launched in June 2026, is led by Adam Selipsky, former CEO of AWS, and counts KKR, the Kuwait Investment Authority, NVIDIA and U.S. power company Vistra as founding investors. Here is the deal structure, what Helix actually builds, why power is the stated bottleneck, and where each part of Samsung's stack fits.

1. The Deal: Who Is Investing, and How Much

Pin the facts first. On September 29, 2026, Samsung Electronics, Samsung C&T, Samsung SDS, Samsung SDI, Samsung Life Insurance and Samsung Fire & Marine Insurance announced a combined USD 1 billion investment in Helix Digital Infrastructure, an AI infrastructure company established by KKR. Samsung Electronics contributes USD 500 million, with the five other affiliates covering the remainder. Helix was launched in June 2026 and is positioned as an AI-enabling infrastructure provider designed to deliver integrated solutions across the entire value chain. Alongside Samsung, KKR, the Kuwait Investment Authority, NVIDIA and U.S. power company Vistra are founding investors. The company is led by Adam Selipsky, former CEO of AWS, and is assembling a management team of experts from the data center and power industries, leveraging KKR's global infrastructure business with about 170 dedicated professionals.

# Helix is not a single asset; it is an integrated chain. The deal is a bet
# that whoever controls data-center AND power capacity controls AI supply.
# Model the investment as exposure across the whole chain, not one layer.

HELIX_STACK = {
    "compute":   ["hyperscale data center development", "data center operations"],
    "power":     ["baseload generation", "flexible energy sources",
                  "transmission & distribution"],
    "network":   ["fiber-optic networks"],
}

def chain_coverage(portfolio: dict) -> dict:
    covered = {layer for layer, parts in portfolio.items() if parts}
    return {
        "covered": covered,
        "missing": set(HELIX_STACK) - covered,
        "thesis": "own the whole chain so capacity cannot be externally throttled",
    }

Data centers turning from facilities into platforms

2. What Helix Actually Builds: One Integrated Chain

Helix's proposition is to treat infrastructure as one integrated chain. The official description covers four segments: hyperscale data center development and operations; power generation, covering both baseload and flexible energy sources; transmission and distribution infrastructure; and fiber-optic networks. Its customers are hyperscalers, who can rapidly secure the large-scale, comprehensive infrastructure required to keep pace with exponentially growing AI demands. Writing the chain out makes it concrete: the compute layer is data center development and operations, the energy layer is generation plus transmission and distribution, and the network layer is fiber. Code sample 1 structures that chain and uses it to test whether a portfolio covers every link; the more links missing, the more capacity can be throttled from outside.

# "If data centers are the production hubs of the AI era, power is the fuel
# that drives them." Power is described as the greatest bottleneck, so size
# the facility to the electrons you can actually secure, not the other way.

MW_PER_RACK = 0.13          # order-of-magnitude for dense accelerated racks

def site_capacity(racks, available_mw, pue=1.3):
    it_load = racks * MW_PER_RACK
    total_load = it_load * pue          # includes cooling overhead
    return {
        "racks": racks,
        "it_mw": round(it_load, 1),
        "total_mw": round(total_load, 1),
        "power_bound": total_load > available_mw,
        "max_racks_by_power": int(available_mw / (MW_PER_RACK * pue)),
    }

# Helix plans to secure energy capacity through direct investments and
# partnerships with major energy developers, including investor Vistra.

3. Why Power Is the Current Bottleneck

This is the sentence worth remembering from the whole deal. Samsung frames data centers as the production hubs of the AI era and power as the fuel that drives them, and states directly that power availability is currently the greatest bottleneck in AI infrastructure. As a result, the ability to secure both compute and power infrastructure at the same time has become a defining competitive advantage. Helix's strategy targets precisely that, planning to secure energy capacity rapidly through direct investments and strategic partnerships with major energy developers, including energy companies such as investor Vistra. What does that mean for engineers? Size capacity planning around the electrons you can secure rather than the rack count. Code sample 2 gives a minimal model that puts racks, IT load, PUE and available megawatts in one place.

# The new data-center business models are why power interconnection timing
# drives the whole deal. GPUaaS, sovereign AI and colocation each price
# differently, and each is gated by energisation, not by silicon availability.

BUSINESS_MODELS = {
    "GPUaaS":        "rent accelerated compute by the hour; margin tracks power cost",
    "Sovereign AI":  "host in-jurisdiction for residency and control; buyer pays premium",
    "Colocation":    "rent space and power; power capacity is the scarce unit",
}

def time_to_revenue(model, months_to_power, months_to_racks):
    energised = max(months_to_power, months_to_racks)
    return {
        "model": model,
        "gated_by": "power" if months_to_power >= months_to_racks else "racks",
        "months_to_revenue": energised,
    }

Power availability is the current bottleneck

4. Where Each Samsung Affiliate Fits

Samsung says the investment lets it move from a traditional hardware component supplier to an active architect of the global AI infrastructure ecosystem. Piece by piece, the roles are specific. Samsung Electronics' DS Division supports the global build-out with advanced semiconductor solutions, responding to shifts in demand in a timely manner. The DX Division provides a comprehensive data center cooling portfolio, from air cooling systems to coolant distribution units for liquid cooling, through FläktGroup, the data center HVAC specialist it acquired in 2025, and operates 14 production sites and a supply and service network across 65 countries. Samsung C&T's Engineering & Construction Group drives large-scale infrastructure projects as an EPC contractor across data centers and power generation. Samsung SDS designs, builds and operates data centers, has entered the GPUaaS business, and holds the lowest power usage effectiveness in Korea, applied to its own data centers and the Korea AI Computing Center. Samsung SDI is recognized for world-leading uninterruptible power supplies and battery backup units. Code sample 4 maps that in-house capability to the risk it retires.

# Samsung's affiliates each own a link in the chain, which is why a single
# group can anchor a build-out: silicon, cooling, EPC, operations, and power
# backup are all in-house. Map vendor capability to the risk it retires.

SAMSUNG_SUPPLY = {
    "Samsung Electronics (DS)": "semiconductors for the compute layer",
    "Samsung Electronics (DX)": "data center cooling (FläktGroup), incl. liquid CDUs",
    "Samsung C&T":              "EPC for data centers and power generation",
    "Samsung SDS":              "data center design/ops + GPUaaS (lowest PUE in Korea)",
    "Samsung SDI":              "UPS and battery backup units (BBU) for 24/7 servers",
}

def de_risk(supply: dict, needed: list) -> dict:
    # A vertically capable group can retire single-vendor and integration risk.
    return {k: ("in-house" if k in supply else "external") for k in needed}

5. The New Business Models and Power-Gated Deployment

Samsung notes that AI data centers are rapidly evolving beyond simple facility investments into core business platforms, integrating high-value models such as GPU-as-a-Service, Sovereign AI and colocation. Together they drive a multi-sector ecosystem spanning power generation, cooling technologies, semiconductors and finance. The key implication for builders is that time-to-value is gated by energisation rather than by silicon arrival: when a campus starts producing revenue often depends on interconnection queues, permitting and fuel contracts more than on GPU lead times. Code sample 3 folds these three models and the relationship between months-to-power and months-to-racks into a simple payback model. Code sample 5 offers a signal function for where the bottleneck is migrating, on the principle that you build and price around the longest active constraint.

# Watch the bottleneck migrate. Silicon shipped on a roadmap; power ships on
# permitting, interconnection queues and fuel contracts. The party that can
# sequence electrons to racks on time wins the deployment, so track it.

def bottleneck(market):
    signals = {
        "silicon": market.get("gpu_lead_time_weeks", 0) > 26,
        "cooling": market.get("liquid_ready_racks_pct", 100) < 60,
        "power":   market.get("interconnect_wait_months", 0) > 24,
        "fiber":   market.get("dark_fiber_available", True) is False,
    }
    active = [k for k, v in signals.items() if v]
    return {"active_constraints": active,
            "rule": "build and price around the longest active constraint"}

Samsung's affiliates each hold a position

6. What This Means for Builders

Finally, the actionable judgements. First, treat power as the primary constraint: when planning capacity, compute the electricity you can secure first and work back to racks and load rather than the other way around, as code sample 2 does. Second, when evaluating suppliers, ask whether they can retire risk across multiple links: silicon, cooling, EPC, operations and power backup coordinated inside one group, as code sample 4 frames. Third, leave architectural room for sovereign AI and data residency requirements, which tend to be premium work. Fourth, build cost models on cost per completed task rather than per token, because power and cooling overhead eventually lands in the inference price. Fifth, watch for the constraint migrating: when interconnection waits exceed 24 months or liquid-ready rack coverage falls below 60 percent, the bottleneck has moved, which code sample 5 encodes. The thesis behind this billion-dollar bet is simple: whoever can sequence the chain from electrons to tokens on time controls AI supply.