Digital Infrastructure Insights October 2026: prefabricated AI factories and reference designs promised deployment in weeks rather than months, power and cooling became the pacing items for new capacity, telcos and national champions built sovereign AI sites from Taiwan to Norway, financing reached new scale while utilisation data raised ROI questions, chip packaging and substrates became the next supply constraint, and fibre and optics for AI interconnect drew major deals.
At a glance: Digital Infrastructure Insights October 2026
- AI factories became products: NVIDIA‘s DSX POD aims to cut deployment from about six months to as little as 14 days, Schneider Electric reported 80% faster onsite installation with prefabricated modules, and HPE, Dell, Cisco and Nutanix expanded packaged AI stacks.
- Power and cooling set the pace: National Grid and NVIDIA reported up to 40% lower cluster power draw in tests, Vertiv completed its $1.45 billion acquisition of UtilityInnovation Group, Carrier is investing more than $600 million in a chiller plant, and Rystad Energy sees data centre fuel cell investment reaching about $30 billion by 2030.
- Sovereign AI capacity spread: Chief Telecom, Bell Canada, Globe Telecom, TIM, Naver and HUMAIN all advanced national AI sites, and Cortex DC plans 2 GW of AI-ready capacity across India.
- Financing scaled while ROI came under scrutiny: Broadcom arranged $63 billion of financing for Anthropic AI chips and SpaceX secured about $40 billion for NVIDIA GPUs, while Gartner estimated a $4.9 trillion revenue gap by 2030 and VentureBeat found 81% of companies use half or less of their GPU capacity.
- Packaging and substrates became the constraint: NVIDIA signed capacity agreements to raise substrate output 2.5x over three years, and GlobalFoundries and TSMC signed a $2 billion silicon interposer deal.
- AI interconnect drew investment: Verizon signed a dark fibre agreement with AWS worth more than $2 billion, Zayo plans more than 8,000 miles of fibre with NVIDIA and Corning, and CBRE agreed to acquire Onnec for up to $1 billion.
1. AI factories become catalogue products
Suppliers moved from bespoke AI builds to pre-integrated, factory-built products this month, with time to first token becoming the headline metric.
- NVIDIA: introduced DSX POD, a factory-built reference design for containerised Vera Rubin AI factories that arrive pre-integrated and validated, reducing deployment from about six months to as little as 14 days, with partner delivery from early 2027; CEO Jensen Huang said a three-month delay in AI data centre infrastructure can forgo roughly $10 billion in revenue.
- Schneider Electric: unveiled its RapidScale Modular Data Center Reference Design 115 for Vera Rubin; a joint study with Compass Datacenters found prefabricated IT modules cut onsite installation time by 80%, from 75.5 hours to 15 hours per pod.
- ZTE: launched the M8000 micro-module with hybrid air-liquid cooling and 800V HVDC power, supporting 20 to 60 kW racks and a PUE as low as 1.15, according to ZTE; Vertiv and NVIDIA aligned virtual validation with factory-built deployment, adopted by Polar Data Centres.
- Packaged stacks: HPE added a liquid-cooled NVIDIA Vera CPU rack with 256 Vera processors and support for 22,500 AI agent sandboxes to Private Cloud AI, Dell advanced enterprise agentic AI from deskside to data centre with NVIDIA, Cisco expanded its Secure AI Factory with NVIDIA and Supermicro, and Nutanix partnered with NVIDIA, Supermicro and Netris on Spectrum-X based infrastructure.
- Operations: NVIDIA launched Minerva, an AIOps platform for AI factory observability with OEM distribution through Dell, Lenovo and Supermicro, and Equinix opened liquid-cooled Enhanced Performance Innovation Centers in Frankfurt, London, Osaka, Paris and Sydney for testing before production.
- Open standards: at the OCP Global Summit, NVIDIA contributed Vera Rubin rack designs supporting up to 250 kW racks, joined AMD, Google and Meta to publish Open Data Center Specification v0.7, and with Google, Meta and Microsoft released a Grid Resilience Specification targeting sub-2-second recovery from grid events.
- CPU capacity for agents: AMD‘s Helion rack-scale architecture supports up to 32,768 sixth-generation EPYC cores, with initial availability targeted for November 2026, and Oracle made NVIDIA Vera Rubin NVL72 and Vera CPUs generally available on OCI.
- TeckNexus analysis: “AI Factories Become Products: The Race to Shorten Time to Token” examines how open standards and integrated stacks are converging.
Why it matters for buyers: prefabricated designs shift risk from the construction site to the factory and make AI capacity easier to compare as a product. Buyers should ask suppliers for time-to-first-token commitments, site readiness requirements and validation evidence, not only rack specifications.
2. Power and cooling set the pace
Electricity access and thermal management determined how fast new AI capacity could come online, and suppliers from turbines to chillers expanded in response.
- National Grid and NVIDIA: reported up to a 40% reduction in cluster power draw during tests on a Blackwell Ultra GPU cluster; NVIDIA’s DSX AI Factory Platform in Europe targets near-zero facility cooling water use and up to 30% lower power for the same token output.
- On-site power: Rystad Energy expects data centre fuel cell investment to rise from $2.8 billion in 2025 to about $30 billion by 2030, with Oracle planning to procure up to 2.8 GW of fuel cell systems; Crusoe secured a $700 million facility from Neuberger Specialty Finance for behind-the-meter power equipment.
- Vertiv: completed its $1.45 billion acquisition of UtilityInnovation Group, adding microgrid controls, on-site generation orchestration and behind-the-meter power architecture.
- Manufacturing capacity: Carrier is investing more than $600 million in a Fort Worth, Texas chiller plant targeting about 3,000 units a year from 2027, and Mitsubishi Power Americas is investing $490 million to expand gas turbine component production in Georgia.
- Grid delays: construction of a data centre in Eagle Mountain, Utah slipped because of limited grid capacity, and Nscale faced delays, equipment failures and power constraints at some sites.
- Longer-term supply: Helion signed a letter of intent for up to $175 million in CHIPS Act funding, alongside $50 million in private investment, for semiconductors that could improve fusion power plant output, and Peak Energy raised $108 million in a round led by GM Ventures.
- Communities: AWS committed more than $1 billion over five years to programmes for US data centre host communities, Google, Meta, AWS and Microsoft increased transparency on data centre plans, and Digital Realty proposed a 1 GW campus in Kokomo, Indiana by 2033 using closed-loop cooling.
- Reuse: NVIDIA is exploring converting shut-down German nuclear power plant sites into AI data centre campuses to reduce build cost and time.
Why it matters for buyers: power availability and cooling equipment lead times now decide project schedules more than GPU supply. Buyers should test site power plans, including on-site generation and grid timelines, and confirm cooling capacity before committing to delivery dates.
3. Sovereign AI capacity: telcos and national champions
Telecom operators and national technology groups turned land, power and fibre into in-country AI capacity, often with government backing.
- Chief Telecom (Taiwan): broke ground on a new AI data centre at the Central Taiwan Science Park, investing about NT$3.3 billion (US$101.7 million), with operations from the second half of 2028.
- Telcos: Bell Canada accelerated an AI data centre near Regina with Saskatchewan government backing, Globe Telecom plans at least $1 billion of data centre investment in the Philippines, and Elisa secured about €100 million in cumulative data centre connectivity orders in the first nine months.
- TIM (Italy): TIM Enterprise will host a sovereign AI platform combining its cloud and edge network with NVIDIA GPUs, Nemotron models and NeMo Guardrails.
- Naver (South Korea): secured an agreement with Brookfield to finance up to $9 billion of AI infrastructure, launching with 55 MW in H1 2027 and expanding to 200 MW by 2028, with NVIDIA planning a $1 billion strategic investment; SK Telecom’s SK Hyper is promoting South Korea as a location for AI model training.
- HUMAIN (Saudi Arabia): is scaling with an 84 MW NVIDIA-based deployment and a 50 MW offtake agreement with Nebius at Oxagon.
- India: Cortex DC plans 2 GW of AI-ready capacity with its first site within about 18 months and 51% renewable energy, and Neysa and STT GDC will deploy up to 9,000 GPUs at STT Noida 2 from Q1 2027.
- Europe and the Americas: Alibaba Cloud opened its first public cloud region in the Netherlands, WWT and Polarise plan a 5 MW sovereign AI factory in Oslo, VivoPower and Open Innovation AI signed a framework targeting at least 1 GW of sovereign AI data centres, and Crusoe and Araico will build a hydro-powered AI cloud site in Paraguay starting at 106 MW of IT capacity.
- TeckNexus analysis: “Telcos as Sovereign AI Landlords: Turning Land, Power and Fibre into National AI Capacity” sets out four sovereign AI infrastructure models for operators.
Why it matters for buyers: sovereign AI options are multiplying, and many come from local telcos with existing fibre and power assets. Regulated buyers can now compare in-country capacity on residency, governance and price, rather than defaulting to the nearest hyperscale region.
4. Financing reaches new scale as ROI comes under scrutiny
October brought some of the largest AI infrastructure financings yet, alongside data suggesting much installed capacity is underused.
- Broadcom: arranged $63 billion of financing with Bank of America, Blackstone, Citigroup and Morgan Stanley for a special-purpose vehicle that will buy its custom XPU accelerators and networking equipment for lease to Anthropic, with deliveries from 2027.
- SpaceX: secured a binding commitment for about $40 billion of financing led by Apollo Global Management for NVIDIA Vera Rubin GPUs on five-year leases, with deliveries from H1 2027; it reported $7.6 billion in AI segment revenue in Q3 2026.
- Neoclouds and platforms: Nebius closed a $3.5 billion facility led by Citi, Argentum AI is reportedly negotiating $5 billion of financing for a 300 MW site in Poland, AM Intelligence is reportedly seeking up to $3 billion for Vera Rubin chips, and DayOne Data Centers filed for a Nasdaq IPO of up to US$5 billion at about a US$20 billion valuation.
- Gartner: estimated that data centre owners will need $8.5 trillion in cumulative revenue by 2030 to cover AI capex and a 12% return hurdle, against about $3.6 trillion attributable, a gap of roughly $4.9 trillion.
- Utilisation: a VentureBeat VB Pulse survey found 81% of companies with AI in production use half or less of their GPU capacity, while 71% still plan new AI infrastructure.
- Costs and pricing: Bank of America estimates AI data centre construction at $49 million per MW, and NVIDIA discussed a possible annual software fee of about $2,000 per GPU for running agents, though pricing and scope are unconfirmed.
- Capital in Asia: Blackstone plans to double India investments to $100 billion over five years with data centres a priority, SC Capital Partners invested in a 40 MW AI data centre in Pohang, while Firmus withdrew a planned IPO and PLDT postponed its data centre REIT listing to 2027.
- TeckNexus analysis: “The AI Infrastructure ROI Reckoning: Utilisation, Speed and Smarter Financing” examines the three levers reshaping returns.
Why it matters for buyers: lease-based and asset-backed structures are making large GPU fleets easier to fund, but low utilisation weakens the case for new capacity. Buyers should model utilisation honestly and compare owning, leasing and consuming capacity as a service before committing.
5. Chips, packaging and supply chain resilience
The constraint on AI hardware moved further down the supply chain this month, to substrates, packaging, testing and critical materials.
- NVIDIA: signed capacity agreements with IBIDEN, Samsung Electro-Mechanics and Unimicron to raise substrate output 2.5x over three years; Amkor will expand interposer packaging capacity in South Korea by more than 300% by end-2028, KYEC will add US testing facilities backed by a $1.1 billion advance, and SPIL targets a 15x increase in co-packaged optics engine capacity by mid-2027.
- GlobalFoundries and TSMC: signed a five-year, $2 billion agreement for silicon interposers made in Malta, New York, ramping in H1 2028; Japan will fund $20 billion for two new GlobalFoundries fabs at the same campus.
- TSMC: is weighing a multibillion-dollar Texas campus and exploring rare-earth-free motors in factory equipment to reduce exposure to export restrictions; Intel installed the first major structural truss at its Ohio fab.
- Memory and materials: SK hynix is developing an iHBM thermal architecture that cuts memory thermal resistance by more than 30%, with HBM5 targeted for 2029, and Infineon will invest about €1.1 billion in a high-volume 300mm GaN power semiconductor fab in Villach, Austria.
- India: Tata Sons invested about $115 million in Tata Electronics, whose Dholera fab is expected to start operations in 2028, while chip design startups report paying about 50% more for wafer fabrication than larger customers.
- Systems supply: HPE will supply AMD Helios AI racks, networking and software to Vultr under a $1.2 billion agreement, Oracle and AMD expanded Helios on OCI, and Foxconn is preparing to expand its AI data centre hardware partnership with OpenAI.
- TeckNexus analysis: “Beyond the GPU: Power and Packaging Are Now the AI Infrastructure Bottlenecks” maps where the AI constraint is moving.
Why it matters for buyers: lead times for AI systems now depend on packaging and substrate capacity as much as on GPU allocation. Buyers should ask suppliers about their position in these supply chains and include delivery risk in supplier scoring.
6. Fibre, optics and networking for AI
AI clusters spread across sites and regions, and the fibre, optics and switching between them attracted major deals.
- Verizon: signed a dark fibre agreement with AWS worth more than $2 billion to support distributed compute for AI workloads, alongside a similar arrangement with Google.
- Zayo: plans more than 8,000 miles of fibre backed by its NVIDIA partnership and a long-term supply agreement with Corning.
- Lumen: is combining Eminence Grey’s sovereign AI platform with its programmable network to connect edge and cloud AI environments.
- Switching and optics: Meta is adopting NVIDIA Spectrum-6 (102.4 Tbps) Ethernet switches, NVIDIA and Cisco unveiled the USFP optical standard targeting 400 Gbps per lane, Cisco expanded AI networking with new silicon and SONiC capabilities, and Semtech secured SiGe capacity at Tower Semiconductor for 800G to 3.2T optical interconnects.
- Marvell: is reportedly in advanced talks to acquire Lumentum, a supplier of optical components for AI data centres; no agreement has been confirmed.
- Services: CBRE agreed to acquire Onnec, a fibre networking and structured cabling specialist, for up to $1 billion, and MRO-TEK Networks is developing an indigenous network operating system for data centre switching in India.
- Edge compute: Extreme Networks plans on-premises AI infrastructure that uses idle compute capacity in network equipment.
Why it matters for buyers: as AI workloads spread across sites, interconnect capacity and optical supply become part of the compute decision. Buyers should plan fibre routes, diversity and optics lead times alongside GPU and power plans.
→ Plan interconnect and capacity between sites with the TeckNexus Network Planning tools.
What Digital Infrastructure Insights for October 2026 means if you’re evaluating AI infrastructure
October’s throughline is speed with discipline: AI capacity can now be ordered like a product, but power, supply chains and returns decide whether it pays. Six moves follow directly from the month:
- Ask for time to first token: DSX POD, Schneider Electric’s prefabricated modules and packaged stacks from HPE, Dell, Cisco and Nutanix make deployment speed a contract term.
- Secure power before hardware: fuel cells, behind-the-meter generation and grid timelines now set the schedule, as Eagle Mountain and Nscale show.
- Compare sovereign options: telcos and national champions in Taiwan, Canada, Italy, South Korea, Saudi Arabia and India now offer in-country AI capacity.
- Model utilisation honestly: with 81% of AI users running GPUs at half capacity or less, test own, lease and as-a-service options before buying.
- Score supply chain risk: substrates, packaging and testing capacity now shape delivery dates, so include them in supplier evaluation.
- Plan interconnect early: dark fibre, optics and switching deals at Verizon, Zayo and CBRE show that links between sites are part of the AI build.
Catching up? Last month’s signal is here: Gigawatts, GPUs and Sovereign Stacks: Digital Infrastructure Insights, September 2026 →
→ Start with the TeckNexus Intelligence Platform: independent, buyer-neutral tools for infrastructure, TCO, architecture, and RFP decisions.
This analysis is drawn from TeckNexus’s full curated Digital Infrastructure Monthly Insights for October 2026. See every deployment, product, and partnership update →
