Who Benefits Most from the $725B DC Capex Wave?

2026 data centre capex from hyperscalers plus neoclouds totals ~$725B, up ~36% YoY. But these companies were already spending heavily, so the real question is: where does the incremental ~$190B of new spending land, who sees the biggest revenue uplift relative to their existing business, and who has the supply constraints to protect margins?

---

Tier 1 — Transformed: DC Capex IS the Company

These businesses are being fundamentally reshaped. DC demand matches or exceeds their entire current revenue.

---

VRT — Vertiv (Cooling & Power Distribution)

$10B revenue → $10-13B DC demand = ~115% of current revenue

Arguably the most transformed company in the entire chain. Five years ago Vertiv was a sleepy $5B industrial business. Now orders are up 60% YoY, backlog is $9.5B with a 1.4x book-to-bill. Every megawatt of DC capacity needs cooling racks and power distribution — and Vertiv is one of very few scaled suppliers. Supply is genuinely constrained, supporting pricing. The incremental uplift from 2024 to 2026 could be 50-70% revenue growth.

ANET — Arista Networks (DC Networking)

$9B revenue → $8-10B DC demand = ~100% of current revenue

Near pure-play DC networking, taking share from Cisco in back-end AI cluster connectivity. AI-specific networking revenue was ~$750M in 2025, targeting $1.5B+ in 2026. The shift to 800G+ Ethernet for AI clusters is a product cycle Arista owns. Every new GPU rack needs high-bandwidth switching — as the installed GPU base doubles, networking demand follows.

NVDA — NVIDIA (GPUs + NVLink + Systems)

$213B revenue (FY26) → ~$200B DC = ~90%. FY27 consensus: $316B

Already transformed — DC revenue grew from $15B (FY23) to ~$188B (FY26) in three years. The story is now about sustaining 50%+ growth, not discovering it. Blackwell and Vera Rubin have half a trillion in visibility. Supply remains constrained (GPUs sold out), keeping gross margins at 73-75%. The long-term risk is custom silicon gradually eroding ~90% market share, but that’s a 2028+ concern.

---

Tier 2 — Major Driver: DC Capex Reshaping 40-65% of Revenue

DC spending is the dominant growth engine but these companies retain meaningful diversification.

TSM — TSMC (Fabricates Everything)

$122B revenue (CY25) → ~$75-90B DC-linked = ~65%. Guiding +30% for CY26

The ultimate picks-and-shovels play. TSMC wins regardless of whether NVIDIA, Broadcom, AMD, or custom silicon wins — they fab all of them. HPC/AI is now 58% of revenue (up from 39% in 2022). Annual price increases locked in through 2029 because there is no alternative at advanced nodes. The supply constraint is physics — building leading-edge fabs takes 2-3 years, so capacity stays tight through 2027+.

AVGO — Broadcom (Custom ASICs + AI Networking)

$64B revenue (incl. $27B VMware) → ~$30-40B DC semi revenue = ~55%

The hidden story is how fast AI is reshaping the semiconductor half. AI revenue was ~$24B in FY2025, growing 60%+ YoY, with a $73B AI order backlog. Broadcom designs the custom ASICs for Google (TPUs), Amazon, and now OpenAI. As hyperscalers diversify away from NVIDIA, Broadcom is the direct beneficiary. AI networking adds another layer. The incremental growth is concentrated in the semi side — VMware software grows at low-double-digits.

MU — Micron (HBM & Server DRAM)

$42B revenue (TTM) → ~$15-20B DC-linked = ~42%. Q1 FY26 was record $13.6B

Every AI GPU needs 6-8 HBM stacks, and HBM is in severe shortage. Micron is the only US-listed pure play — SK Hynix (~50% share) and Samsung (~25%) are Korean-listed. HBM pricing runs at 3-5x regular DRAM per bit, and the constraint won’t ease before 2027. Margins went from negative to record levels in 18 months. Risk: Korean competitors and eventual oversupply could compress pricing.

---

Tier 3 — Solid Tailwind: DC is 10-38% of Revenue

Meaningful exposure but well-diversified. Less upside but also less risk if DC spending slows.

PWR Quanta Services / EME EMCOR — DC Construction

PWR: $28B rev, ~$8-12B DC (~36%). EME: $16B rev, ~$5-7B DC (~38%)

Record backlogs ($39B for PWR) driven by DC buildout AND the grid infrastructure to power it. The key constraint is skilled craft labour — you can’t quickly train data centre electricians, which protects margins. They benefit twice: building the DCs and building the transmission/substation infrastructure to feed them. Lower risk than pure-play DC names because they also serve grid, renewables, and industrial markets.

ETN Eaton / CAT Caterpillar — Power Infrastructure

ETN: $27B rev, ~$6-9B DC (~28%). CAT: $66B rev, ~$5-8B DC (~10%)

Eaton (UPS, switchgear, PDUs) has more DC leverage at ~28% of revenue, with 12-18 month lead times supporting pricing. Caterpillar (backup generators) benefits but DC is ~10% of a massive diversified business — a nice tailwind, not thesis-changing. Both benefit from supply constraints on electrical equipment but neither is being reshaped by this cycle.

---

Why Supply Constraints Matter: Pricing Power Across the Chain

The revenue uplift is only half the story. What makes this cycle particularly attractive is that supply bottlenecks are creating a seller’s market at almost every layer, meaning growth comes with margin expansion rather than compression.

TSMC — monopoly on advanced nodes, annual price rises locked in through 2029. NVIDIA — GPUs sold out, 73-75% gross margins at massive scale. HBM memory — physical shortage, 3-5x premium pricing. Vertiv/Eaton — cooling and power equipment backlogs at record levels, 12-18 month lead times. Quanta/EMCOR — skilled labour shortage in electrical trades creates barriers to competition.

These constraints look durable through at least 2027.

---

What to Watch

Custom silicon adoption — if hyperscalers shift faster to Trainium/TPUs/Broadcom ASICs, NVIDIA’s share erodes but TSMC and AVGO benefit. Net neutral for the overall pool.

HBM supply normalisation — if the memory shortage resolves, HBM pricing compresses and Micron’s margins come under pressure. Likely a 2027-2028 event.

Neocloud financial health — CoreWeave has $14B debt with $9.7B maturing within 12 months. If neoclouds struggle, ~$55B of the capex pool is at risk.

Capex sustainability — hyperscalers now spend 45-57% of revenue on capex. Any pullback in AI demand or ROI questions could trigger rapid capex cuts, hitting all suppliers simultaneously.

---

Sources: Company earnings reports (Q3/Q4 2025), Dell’Oro Group, IoT Analytics DC Infrastructure Report, IEEE ComSoc, analyst estimates. Revenue figures are latest reported full-year or TTM. DC revenue estimates are approximate allocations from the ~$725B total pool based on industry breakdowns and company disclosures. February 2026.

The market keeps talking about an “AI bubble.” That misses the real fragility.

The bigger bubble risk is in SaaS application software—because AI enables new entrants to build broader, business-native solutions that solve conjoined problems end-to-end, not isolated tasks. That shifts value away from “apps glued together with humans” toward platforms that understand the business process.

The Nokia Trap (what’s actually changing)

Nokia didn’t lose because people wanted a cheaper phone or a slightly better phone.

It lost because the category stopped being “a phone.” The product became a computing platform that absorbed whole new jobs (apps, internet, maps, media). Nokia was optimised for the old definition of the product.

SaaS is facing the same trap:

• Many SaaS tools are optimised for problems not productivity of work
• The next generation solves the whole job, across steps, systems, and teams

When the category expands, the old leaders can look suddenly narrow.

The real threat: “bigger problem” products

Most businesses don’t have “a CRM problem” or “a ticketing problem” or “a billing problem.”

They have a revenue problem, a service problem, a fulfilment problem, a compliance problem—spanning multiple tools and handoffs.

Classic SaaS often works like this:

• a set of apps
• connected by integrations
• held together by people doing manual work, chasing exceptions, updating fields, and reconciling gaps

AI-native challengers can compete on a different axis:

• one model of the business workflow
• one place where context lives
• automation that moves work forward across steps
• handling exceptions without a human being the glue

That’s not “cheaper seats.” It’s a different product category: outcome-focused systems rather than tool-focused apps.

Why this hits SaaS valuations

If buyers shift from “best tool” to “best end-to-end outcome,” the SaaS moat changes:

• Feature depth matters less than workflow ownership
• Integrations matter less than shared context
• UI matters less than automation and decisioning
• Point solutions get re-framed as “components,” not “systems”

That’s exactly where multiples compress: when a market realises a product is becoming a replaceable module inside a bigger platform.

What to watch (the early signs)

Not “seat price down.” Look for:

• new vendors winning by promising cycle-time reduction (close faster, ship faster, resolve faster)
• incumbents expanding into adjacencies defensively (copying platform moves)
• customers consolidating around process owners (systems of action), not systems of record
• services headcount shifting from “ops glue” to “exception oversight”

Bottom line

AI isn’t just making software cheaper.

It’s enabling a change in what customers buy: solutions that map to the business, solve multiple connected problems, and reduce the need for humans to stitch workflows together.

That’s the Nokia Trap for SaaS: people keep asking for “better apps” right up until the product category shifts to “business platforms that run the work.”

“–” More exposed SaaS / app-layer

• Salesforce (CRM) — CRM
• ServiceNow — NOW
• Workday — WDAY
• Adobe — ADBE
• Atlassian — TEAM
• Snowflake — SNOW
• Datadog — DDOG
• HubSpot — HUBS
• Zoom — ZM
• DocuSign — DOCU
• Okta — OKTA
• Twilio — TWLO
• Shopify — SHOP
• Monday.com — MNDY
• Asana — ASAN
• Smartsheet — SMAR
• PagerDuty — PD
• Freshworks — FRSH
• Elastic — ESTC
• GitLab — GTLB
• MongoDB — MDB
• Cloudflare — NET
• Unity — U
• Toast — TOST

“+” Likely beneficiaries (AI platform / infra / “system-of-action” enablers)

• Palantir — PLTR
• Microsoft — MSFT
• Alphabet — GOOGL
• Amazon — AMZN
• Meta — META
• NVIDIA — NVDA
• AMD — AMD
• Broadcom — AVGO
• TSMC — TSM
• ASML — ASML
• Arista Networks — ANET
• Super Micro Computer — SMCI
• Oracle — ORCL
• IBM — IBM
• Cisco — CSCO
• Palo Alto Networks — PANW
• CrowdStrike — CRWD
• Zscaler — ZS

3 min read | For investors tracking NVDA, hyperscalers, and AI infrastructure

---

Emerging AI architectures — particularly Mamba and State Space Models (SSMs) — threaten to disrupt the current AI landscape.

Both hardware and software leaders face risk, though hardware companies are more exposed: custom silicon takes 3-5 years to design, while model companies can retrain and pivot in 12-18 months.

---

The Game-Changer: Mamba and State Space Models

The transformer problem: Today’s AI models (GPT, Claude, Gemini) use “attention” — comparing every word to every other word. Double the context length, quadruple the compute.

That’s why long-context models are so expensive and why hyperscalers are spending hundreds of billions on GPU clusters.

The SSM solution: State Space Models process sequences like a running summary rather than looking at everything simultaneously.

Mamba — the leading SSM architecture, developed by researchers at Carnegie Mellon and Princeton — scales linearly with context length.

The impact:

✓ 5x faster on long sequences

✓ Fraction of the memory footprint

✓ Dramatically lower compute per token

Who’s backing it: Mistral (Codestral Mamba), AI21 Labs (Jamba), Together AI. Hybrid transformer-Mamba models are already in production.

The catch: Pure Mamba hasn’t matched transformers for complex reasoning at frontier scale — yet. But hybrids are proliferating, and the trajectory is clear.

---

What This Means for GPU Demand

If SSMs gain traction, the compute economics of AI change fundamentally.

Transformer world:

• Compute scales quadratically with context
• High memory per token
• GPU hours: baseline
• Long context: expensive, constrained

SSM world:

• Compute scales linearly with context
• Low memory per token
• GPU hours: potentially 50-80% less
• Long context: cheap, abundant

The demand question: Efficiency gains could reduce absolute GPU demand — or unlock new use cases that consume the savings. History suggests the latter, but the transition creates uncertainty.

The NVIDIA risk: SSMs still run on CUDA — but if you need 50-80% fewer GPUs per workload, that’s a volume problem even if NVIDIA keeps 100% share.

The $600 billion hyperscaler capex trajectory assumes insatiable compute demand. SSMs challenge that assumption.

The custom silicon risk: TPUs, Trainium, and MTIA are optimised for transformer attention patterns. SSMs have different compute profiles.

Billions in chips could become sub-optimal — and reduced compute demand means less need for any custom silicon.

---

Who’s Exposed, Who’s Protected

NVIDIA — Moderate exposure

Keeps architecture flexibility, but volume at risk if compute demand drops. Pricing power may erode.

Google, Amazon — High exposure

Custom silicon optimised for wrong architecture and potentially overbuilt for reduced demand. Double hit.

Microsoft — Hedged

NVIDIA dependency means less stranded silicon, but Azure capex still at risk.

Neo-Clouds (CRWV, NBIS) — High exposure

Leveraged to GPU demand. If demand drops 50%, the debt doesn’t.

OpenAI, Anthropic — Lower exposure

Compute costs drop. They’ll adopt what works. Net beneficiaries.

---

The Bottom Line

SSMs and Mamba may not kill transformers outright — hybrids are more likely.

But they rewrite the compute economics and challenge the “insatiable demand” narrative underlying current valuations.

NVIDIA keeps architectural flexibility but faces volume risk. Hyperscalers face both wrong-architecture risk and overbuilding risk. Neo-clouds are leveraged to a demand curve that may flatten.

The model companies may be the cleanest winners — their compute bills go down while their capabilities go up.

---

When efficiency improves dramatically, the biggest spenders have the most to lose.

The AI hardware market presents compelling private investment opportunities as the industry shifts from training to inference. We analysed 12 private companies across accelerators, memory subsystems, and interconnect – here’s what we found.

---

The Big Picture

Four key themes emerged from our analysis:

Inference is where the money is. Training happens once; inference runs continuously at scale. Companies solving inference bottlenecks command premium valuations.

Memory is the constraint, not compute. GPU memory bandwidth limits LLM performance. CXL pooling and in-memory compute bypass expensive HBM.

NVLink lock-in creates openings. NVIDIA’s proprietary interconnect frustrates multi-vendor deployments. Open standards are gaining traction.

Valuations are wildly dispersed. Cerebras at $22B vs d-Matrix at $2B for comparable inference positioning. Early-stage offers better risk/reward.

---

Our Top 5 Picks

Ranked by risk-adjusted return potential for growth-oriented investors.

---

The Rest of the Field (1 = best)

Companies we evaluated but rank lower due to valuation, risk profile, or structural issues:

---

Not Investable

These companies came up in our research but are no longer private opportunities:

• Groq – Acquired by NVIDIA for ~$20B (December 2025)
• Enfabrica – NVIDIA acqui-hired CEO and team for ~$900M (September 2025)
• Rivos – Meta acquisition announced (September 2025)
• Astera Labs – IPO’d March 2024, now $23B+ market cap (NASDAQ: ALAB)

---

Key Risks

Before deploying capital, consider:

⚠️ NVIDIA isn’t standing still. Blackwell and Rubin will improve inference. Startups need to stay ahead of a well-funded incumbent.

⚠️ Hyperscalers are building captive silicon. Google (TPU), Amazon (Trainium), Microsoft (Maia) reduce addressable market.

⚠️ TSMC concentration. Nearly all targets fab at TSMC. Taiwan risk affects the entire portfolio.

⚠️ CXL timing uncertainty. Memory pooling benefits require ecosystem maturity that may take longer than expected.

---

Consider

For a growth mandate targeting 5x+ returns:

---

Next Steps

1. DD on d-Matrix and FuriosaAI
2. Conduct technical DD on Panmnesia’s CXL 3.1 IP
3. Monitor Upscale AI for first silicon tape-out
4. Watch Cerebras IPO for public market entry if pricing rationalises

---

Analysis based on publicly available information as of January 2026. This is not investment advice. Due diligence recommended before any investment decisions.

“`