Pulse (db39922f-26d4-479f-aad7-f443d18dc8be)

Weekly AI infrastructure build-out update: through June 15, 2026

The clean read this week:

> The AI infrastructure trade is turning into a financing, power-access, optics, and full-stack deployment trade. Chips still matter, but the scarce asset is now deployable capacity: compute that can be financed, packaged, powered, cooled, networked, staffed, and contracted.

The biggest change versus prior weeks is that private capital and industrial partners are now stitching the stack together. Apollo, Blackstone, Broadcom, Anthropic, Schneider, Foxconn, Jabil, Adani, Reliance, Meta, Oracle, Applied Digital, Amazon, and Corning all showed up in one week. That is not coincidence. That is the AI build-out becoming an industrial supply chain.

---

1) What changed this week

A) Compute supply chain

Nvidia is tapping debt markets again. Nvidia reportedly plans to raise $20 billion in its first corporate bond sale in five years, with proceeds for general corporate purposes, including refinancing. The larger signal is that AI leaders are increasingly using balance sheets and capital markets to fund the infrastructure arms race, not just operating cash flow. Reuters also noted broader AI-related investment is expected to exceed $700 billion in 2026, up from about $400 billion in 2025.

Broadcom custom silicon got a major validation point. Apollo and Blackstone are backing a $35 billion AI capacity expansion tied to Anthropic, Broadcom custom chips, and networking infrastructure. The project is expected to start with 1 GW of compute capacity in mid-2026 at Fluidstack-managed sites, with a longer-term target above 20 GW by 2028. This is one of the clearest public signals that private capital is willing to finance non-Nvidia AI silicon capacity at massive scale.

Optics moved from derivative theme to supply-chain risk. China’s control over indium phosphide exports is now a real AI data center bottleneck. InP is critical for high-speed photonic chips used in optical links, and China produces roughly 70% of global indium. Reuters cited disruption risk for U.S. photonics names including Coherent, Lumentum, and AXT. Translation: the optical layer is now geopolitically exposed, not just technically constrained.

HBM remains the strategic memory choke point. SK Hynix is reportedly planning a Nasdaq listing, potentially as early as August, after its AI memory-driven rally pushed the company above $1 trillion in market value in May. The listing plan is a market-structure signal: investors want direct exposure to the HBM bottleneck, and SK Hynix wants a broader U.S. AI investor base.

Amazon reinforced the fiber and optical connectivity theme. Amazon signed a multibillion-dollar deal with Corning to expand U.S. optical fiber and connectivity production for data centers. That is a very direct confirmation that data movement is now a procurement priority, not a back-office telecom issue.

---

B) Physical infrastructure

Schneider and Foxconn are productizing AI data centers. Schneider Electric and Foxconn announced a partnership to build ready-to-deploy AI data center infrastructure, combining Foxconn’s manufacturing and AI system capabilities with Schneider’s power, cooling, and energy-management stack. Production is expected to begin later in 2026. This is a major “rack-to-building” integration signal.

Jabil and Adani are building an India AI infrastructure manufacturing platform. Jabil and Adani Enterprises announced a partnership to develop an integrated AI and data center infrastructure manufacturing platform in India. India is now becoming both a demand center and a manufacturing base for AI infrastructure.

Meta and Reliance deepened the India data center theme. Meta agreed to lease an AI-ready data center Reliance will build in Jamnagar. Reuters reported the facility is expected to be 168 MW and partly powered by Reliance renewable assets. This is a concrete example of hyperscalers seeking locations with cheap energy, industrial land, and vertically integrated local partners.

Applied Digital signed another large AI factory lease. Applied Digital signed a 15-year, $5.2 billion lease with an investment-grade U.S. hyperscaler for 210 MW at its Delta Forge 2 AI Factory campus. The company now claims five campuses totaling 1.4 GW of IT load and 2.15 GW of grid-connected power, with committed base-term lease revenue of $36 billion. That is the “power-first data center” model in public-market form.

Oracle went full throttle on AI capex, and investors flinched. Oracle said fiscal 2027 capex could reach $95 billion, well above Wall Street’s expected $67.66 billion, while also planning nearly $40 billion of debt and equity financing. This supports the AI infrastructure bull case, but it also sharpens the bear case: the build-out is becoming debt-heavy, margin-dilutive, and brutally capital intensive.

Power demand keeps moving higher. The U.S. EIA now expects U.S. electricity consumption to hit record highs in 2026 and 2027, rising from 4,195 billion kWh in 2025 to 4,271 billion kWh in 2026 and 4,397 billion kWh in 2027, with AI data centers called out as a key driver. Commercial power use is expected to exceed residential use for the first time on record in 2026.

Gas turbines are now an AI bottleneck too. Siemens Energy said U.S. data centers now account for about 25% of its gas turbine order backlog, while the global market totals around 100 to 120 GW. The company says it has maxed out short-term turbine production capacity and will expand gradually through 2030. In plain English: data centers are now competing with countries, utilities, and industrial buyers for dispatchable power equipment.

---

2) Capex and order trends

Layer This week’s signal Read-through

AI accelerators and custom ASICs Broadcom-linked Anthropic platform backed by Apollo and Blackstone targets 1 GW initially and over 20 GW long term Custom silicon is becoming financeable infrastructure, not just a chip-design project.

Hyperscaler capex Oracle guided fiscal 2027 capex up to $95B AI cloud capex is still accelerating, but debt and free cash flow pressure are now front and center.

Fiber and optics Amazon-Corning multibillion-dollar fiber deal Optical connectivity is becoming a critical AI supply chain.

Photonics materials China InP controls threaten optical chip supply Material supply risk is moving into the AI networking stack.

Data center leases Applied Digital signed $5.2B, 210 MW lease Power-secured AI campuses are getting long-duration contracts.

Power and cooling systems Schneider-Foxconn partnership targets ready-to-deploy AI data centers Infrastructure vendors are moving toward turnkey AI factory platforms.

Power generation Siemens Energy says data centers are 25% of U.S. turbine backlog Dispatchable generation is becoming an AI bottleneck layer.

Grid demand EIA expects record U.S. power consumption in 2026 and 2027 Utility load growth is now structurally linked to AI and electrification.

---

3) Bottlenecks, ranked

Compute bottlenecks

1. HBM and memory bandwidth

HBM remains the highest-value semiconductor choke point. SK Hynix’s U.S. listing plan confirms investor demand for direct exposure to AI memory scarcity.

2. Advanced packaging

CoWoS, 2.5D, 3D packaging, hybrid bonding, substrates, and OSAT capacity remain the toll booths between wafer starts and usable AI accelerators. TSMC has previously said CoWoS capacity CAGR is expected to exceed 80% from 2022 to 2027, which tells you how violent the demand curve is.

3. Optics and photonics materials

The InP story is important because it hits optical chips, not GPUs directly. AI clusters are becoming bandwidth-bound, and the optical supply chain is now a geopolitics problem.

4. Custom ASIC execution

Broadcom is gaining strategic validation, but custom silicon remains concentrated among hyperscalers and a few well-funded AI labs. The winners will be firms with long-term design wins, networking attach, and packaging access.

5. Rack-scale integration

Power shelves, thermal systems, optics, switches, firmware, deployment services, and manufacturing capacity now matter almost as much as the accelerator itself. Schneider-Foxconn is a clean example of this shift.

Physical infrastructure bottlenecks

1. Energized power capacity

The best data center site is not the cheapest land. It is the site with contracted power, grid access, cooling, fiber, and permitting.

2. Gas turbines and dispatchable generation

Siemens Energy’s backlog commentary shows that gas turbines are now in the AI scarcity stack. Data centers want firm power, not vibes and a solar brochure.

3. Transformers, switchgear, substations

Still the ugliest bottleneck. Nobody wants to pitch “medium-voltage switchgear” at a cocktail party, but without it, the AI factory is just an expensive warehouse.

4. Transmission and interconnection queues

Grid connection timing is becoming a competitive advantage for data center developers and utilities.

5. Cooling and water

Applied Digital highlighting waterless cooling is not marketing fluff. Cooling is becoming a permitting, density, and operational-cost differentiator.

6. Construction labor and technicians

Meta’s $115 million investment into data center technician training is a tell: the bottleneck is not only chips and power. It is also people who can build and operate the facilities.

7. Local opposition and ratepayer politics

Reuters/Ipsos found only 14% of respondents were comfortable with a data center being built nearby, while 77% worried AI data centers would raise electricity costs. This is no longer a niche permitting issue. It is becoming political risk.

---

4) Competitive landscape and ticker map

Compute supply chain

Theme Public names

Core GPUs and accelerators NVDA, AMD

Custom ASICs and AI networking silicon AVGO, MRVL, GOOGL, AMZN, META, MSFT, INTC

Foundry TSM, Samsung Electronics, INTC

HBM and memory MU, SK Hynix, Samsung Electronics

Advanced packaging and OSAT TSM, AMKR, ASE

Semi equipment ASML, AMAT, LRCX, KLAC, TEL

Substrates and materials Ibiden, Unimicron, Shinko, Ajinomoto

AI servers and rack systems DELL, HPE, SMCI, Quanta, Wistron, Wiwynn, Foxconn

Switching and Ethernet ANET, CSCO, AVGO, NVDA, MRVL

Optics, photonics, fiber GLW, COHR, LITE, CIEN, CRDO, FN, AXT, STM

EMS and infrastructure manufacturing JBL, FLEX, Foxconn

Physical infrastructure

Theme Public names

Data center operators and REITs EQIX, DLR, IRM, APLD, CORZ, HUT

Digital infrastructure capital DBRG, BAM/BN, BX, KKR, APO, BLK

Electrical equipment ETN, Schneider Electric, ABB, Siemens Energy, GEV, HUBB, POWL

Grid EPC and transmission construction PWR, MTZ, MYRG

Cooling and thermal management VRT, MOD, TT, CARR, DOV, Schneider Electric

Power generation and merchant power CEG, VST, NRG, GEV, Siemens Energy

Regulated utilities NEE, DUK, SO, AEP, ETR, D

Gas engines and backup power CAT, CMI, INNIO, GEV

Fiber and towers GLW, LUMN, FYBR, AMT, CCI, SBAC

Construction/logistics second-order URI, CAT, EXP, MLM, VMC

---

5) Investable themes with the best signal this week

1) Private-capital financed AI compute

Names: AVGO, APO, BX, BLK, KKR, DBRG, BAM/BN

The Apollo-Blackstone-Broadcom-Anthropic structure is the template: specialist capital funds compute capacity, chip vendors provide custom silicon and networking, and AI labs rent the output. This could become the dominant model for labs that cannot or do not want to self-fund hyperscaler-scale infrastructure.

2) Optical connectivity and photonics materials

Names: GLW, COHR, LITE, CIEN, CRDO, FN, AXT, STM

Amazon-Corning validates fiber demand, while the InP export story validates the fragility of the optical chip supply chain. This is one of the cleanest “next bottleneck” themes.

3) HBM scarcity

Names: MU, SK Hynix, Samsung Electronics

This remains the memory choke point. SK Hynix is the cleanest leader, Micron is the U.S.-listed scarcity play, and Samsung is the qualification/turnaround play. The problem is not thesis quality. The problem is valuation after the move.

4) Turnkey AI data center infrastructure

Names: Schneider Electric, Foxconn, JBL, FLEX, VRT, ETN, DELL, HPE, SMCI

Schneider-Foxconn and Jabil-Adani show that the industry is moving toward manufactured, repeatable AI factory modules. This benefits firms that can combine power, cooling, racks, manufacturing, and deployment logistics.

5) Power-first data center platforms

Names: APLD, EQIX, DLR, IRM, DBRG, BAM/BN

Applied Digital’s new lease shows that contracted power and grid-connected sites can be monetized into long-duration hyperscaler leases. This is attractive, but execution risk is high: financing, construction, customer concentration, and power delivery all matter.

6) Electrical grid and power equipment

Names: ETN, GEV, Siemens Energy, ABB, Schneider Electric, HUBB, POWL, PWR

This remains the strongest “boring industrials” basket. It has order visibility, pricing power, and bottleneck status. The catch is that many names are no longer cheap, so pullbacks matter.

7) Gas turbines and dispatchable generation

Names: Siemens Energy, GEV, CAT, CMI, CEG, VST, NRG

Siemens Energy saying U.S. data centers are roughly a quarter of its gas turbine backlog is a giant flare. AI does not just want electricity. It wants firm, available, schedulable electricity.

8) India AI infrastructure build-out

Names: Reliance, Adani Enterprises, JBL, Schneider Electric, Meta indirectly, local power and EPC plays

Meta-Reliance and Jabil-Adani both landed this week. India is now a serious AI infrastructure geography because it offers scale, industrial partners, manufacturing ambition, and power-development optionality.

---

6) My trade read

The AI infrastructure market is splitting into three investable buckets:

Bucket 1: Scarce silicon inputs

HBM, advanced packaging, substrates, ASIC design, optics, and photonics materials. Best tickers: MU, SK Hynix, TSM, AMKR, ASE, AVGO, MRVL, GLW, COHR, LITE, CRDO.

Bucket 2: Deployable systems

Rack-scale servers, power and cooling modules, EMS, and turnkey data center infrastructure. Best tickers: DELL, HPE, SMCI, VRT, ETN, Schneider, JBL, Foxconn.

Bucket 3: Power and real assets

Power-secured data centers, utilities, gas turbines, transmission EPCs, and infrastructure capital. Best tickers: APLD, EQIX, DLR, IRM, GEV, Siemens Energy, CEG, VST, PWR, APO, BX, BAM/BN, BLK.

The highest-signal update this week is that capital formation has become part of the technology stack. Oracle needs financing. Nvidia is tapping bonds. Anthropic capacity is getting backed by Apollo and Blackstone. Applied Digital is monetizing power-backed leases. Schneider and Foxconn are manufacturing deployable AI data center systems.

Bottom line: the best AI infrastructure question is no longer “who has the fastest chip?” It is:

> Who controls a scarce layer that prevents AI capacity from coming online?

Right now, the answer is HBM, packaging, optics, grid equipment, dispatchable power, cooling, and power-secured real estate.

Pulse

Snapshot

Sentiment Read-Through

Run history