Compute & Infrastructure
Top Line
Cerebras Systems raised $5.55 billion in what is 2026's largest IPO to date, signalling that capital markets remain wide open for alternative AI chip architectures that challenge NVIDIA's dominance in inference workloads.
Blackstone executed a dual-track infrastructure capital raise — a $1.75 billion REIT IPO and a $2.3 billion AirTrunk loan for Malaysia expansion — demonstrating that institutional appetite for AI data centre assets is being met with increasingly sophisticated financing structures.
NHN brought a 7,656-GPU cluster online in Seoul, a confirmed capacity addition that reinforces South Korea's push for sovereign AI compute at scale.
SSD prices in Japan have surged up to 300%, with 8TB enterprise drives reaching $3,500, exposing a storage supply crunch that is a direct consequence of AI infrastructure build-out consuming NAND capacity.
Fractile closed a $220 million Series B to develop dedicated AI inference silicon, adding to a growing field of NVIDIA challengers backed by tier-one venture capital.
Key Developments
Cerebras and Fractile Capital Raises Signal Structural Shift in AI Chip Competition
Cerebras Systems priced its IPO at $185 per share, raising $5.55 billion — the largest public offering of 2026 — according to Bloomberg. The company's wafer-scale engine architecture is purpose-built for large-model inference, positioning it directly in the highest-growth segment of AI compute demand. The IPO validates that public markets will price alternative chip architectures at a premium if they can credibly address the inference bottleneck.
Separately, Fractile raised a $220 million Series B co-led by Accel, Factorial Funds, and Founders Fund to accelerate its own inference chip development, per Data Center Dynamics. Both raises reflect the same underlying thesis: inference is where AI compute spend is concentrating as training runs mature, and NVIDIA's H100/B200 stack — designed primarily around training economics — is increasingly perceived as over-specified and over-priced for inference-only deployments. The risk for both companies is that NVIDIA's NIM inference software stack and its roadmap to more inference-optimised products narrow the addressable gap before either reaches volume production.
Blackstone's Dual Capital Raise Accelerates Southeast Asian Data Centre Buildout
Blackstone Digital Infrastructure Trust raised $1.75 billion in its US IPO, with proceeds designated for data centre acquisitions, while simultaneously AirTrunk — also Blackstone-owned — is marketing a $2.3 billion loan specifically for a Malaysia expansion project, per Bloomberg and Bloomberg. The combined capital deployment of over $4 billion from a single infrastructure manager in one week illustrates how private equity is institutionalising AI infrastructure at scale.
Malaysia has become a focal point for Southeast Asian AI capacity for several reasons: competitive land and power costs, a government actively courting hyperscaler investment, and proximity to Singapore's constrained market where a moratorium on new data centre construction has periodically limited supply. The AirTrunk loan is a confirmed financing process — it is being marketed to lenders — but the Malaysia project itself remains a development-stage asset. The critical infrastructure risk is power: Malaysia's grid infrastructure in key corridors is under increasing strain, and securing long-term power purchase agreements at scale is the primary execution risk for projects of this size.
NHN's Seoul GPU Cluster Confirms South Korea as Credible Sovereign AI Compute Node
NHN has brought a 7,656-GPU cluster online at its Yangpyeong-dong facility in Seoul, per Data Center Dynamics. This is a confirmed operational deployment, not an announced plan. At that GPU count, the cluster is sized to support both large-scale model training and high-throughput inference, positioning NHN as a domestic cloud AI provider capable of competing with US hyperscaler footprints in the Korean market.
South Korea's investment in sovereign compute capacity sits within a broader regional pattern — Japan, India, the UAE, and France have all made material commitments to domestically controlled GPU infrastructure in the past 18 months. For South Korea specifically, the motivation is a combination of industrial policy (protecting its semiconductor and technology sector), data sovereignty requirements for regulated industries, and strategic hedging against US export control regimes that have already curtailed chip availability in adjacent markets. The NHN cluster almost certainly relies on NVIDIA hardware, meaning it is sovereign in terms of operation and data control but not in terms of the underlying silicon supply chain.
AI Storage Crunch Surfaces as a Material Supply Chain Vulnerability
Enterprise SSD prices in Japan have surged up to 300%, with Samsung's 8TB 9100 Pro reaching approximately $3,500 at retail, according to Tom's Hardware. The proximate cause is AI infrastructure build-out absorbing enterprise NAND capacity at a rate that has outpaced supply expansion, combined with yen weakness amplifying import costs in Japan specifically.
This is a signal of a broader infrastructure constraint that has received less attention than GPU shortages: AI training and inference pipelines require extremely high-throughput, low-latency local storage, and the enterprise NAND market is not structured to absorb a step-change in demand of the kind AI infrastructure represents. The storage crunch is unlikely to resolve quickly — NAND fab capacity expansions have multi-year lead times, and AI workloads are structurally different from consumer storage demand, requiring different form factors and endurance ratings. Data centre operators planning storage-intensive AI deployments should treat current spot pricing as a constraint on capacity planning, not an anomaly.
Signals & Trends
Optical Interconnects Are Becoming the Next Hardware Chokepoint in AI Data Centres
As GPU cluster density increases and PCIe bandwidth requirements push toward 128 GT/s, the internal networking fabric of AI data centres is becoming a critical bottleneck. Analysis from Semiconductor Engineering highlights that III-V semiconductor lasers — sourced primarily from a small number of specialised compound semiconductor fabs — are essential components for the optical interconnects that make ultra-high-bandwidth GPU-to-GPU communication possible. The supply chain for III-V semiconductors is far more concentrated and less geopolitically diversified than silicon. Meanwhile, PCIe controller architectures are undergoing a structural transition: multistream configurations are shifting from an optional optimisation to a baseline requirement at 128 GT/s, per Semiconductor Engineering. Infrastructure planners who are focused exclusively on GPU procurement are underweighting the interconnect layer as a constraint on cluster performance. The risk is that a facility with next-generation GPUs but inadequate optical interconnect capacity will fail to achieve the theoretical compute density it was designed for.
Financial Structuring for AI Infrastructure Is Outpacing Physical Build Capacity
The volume of capital being raised and deployed for AI data centre assets — Blackstone's $4+ billion in one week, Cerebras's $5.55 billion IPO, Fractile's $220 million raise — is striking relative to the physical constraints on actual construction. Power grid interconnection queues in the US, UK, and Europe run two to four years; hyperscale data centre construction timelines are 18-36 months in best-case scenarios; and cooling equipment supply chains remain constrained. There is a material risk that the financial infrastructure for AI compute is being built significantly faster than the physical infrastructure, creating a gap between committed capital and deliverable capacity. This dynamic typically resolves through one of two mechanisms: cost inflation as competing projects bid for the same constrained resources (power, steel, cooling equipment, skilled construction labour), or asset write-downs when committed timelines slip. Infrastructure professionals should treat aggressive developer timelines with scepticism and pressure-test power procurement assumptions before lender or investor commitments are finalised.
Cisco's AI Pivot Signals Network Infrastructure Is Entering Its AI Build-Out Demand Cycle
Cisco's better-than-expected revenue forecast, accompanied by a restructuring toward AI-focused networking products, per Bloomberg, is an early indicator that the AI infrastructure build-out is now propagating into the networking layer of the stack. GPU clusters and data centre shells attract the headlines, but at scale, AI workloads impose severe demands on switching fabric, load balancing, and east-west traffic management. Cisco's strategic repositioning — combined with the optical interconnect developments noted above — suggests that network infrastructure vendors are beginning to see the AI-driven demand signal that GPU vendors have been experiencing for two years. This is worth tracking as a leading indicator: if networking equipment suppliers begin to show supply constraints similar to those seen in GPUs in 2023-2024, total data centre deployment timelines will lengthen further.
Explore Other Categories
Read detailed analysis in other strategic domains