AI Rack Limits Exposed: Cooling, 800VDC, and Fiber at Data Center Insights 2026 Today

Type: News Repost

AI data center cooling decisions made in 2026 carry 10-to-15-year consequences. Data Center Insights 2026, a two-day virtual conference opening July 15, puts practitioners on record about liquid cooling chemistry, 800VDC power architecture for megawatt racks, and 800G fiber…

> Republished by AIC Engineering. All rights belong to the original publisher; see Source below.

Data Center Insights 2026 opens today at 11:00 AM ET — a two-day virtual conference that arrives at the exact moment data center engineers are being forced to make cooling, power, and fiber decisions with 10-to-15-year consequences. Produced by Endeavor B2B's Data Center Frontier, Cabling Installation & Maintenance, ISE, Lightwave, and SecurityInfoWatch, the event runs through July 16 and draws operators, architects, and engineers navigating what may be the most consequential infrastructure inflection point since the shift to cloud computing.

The urgency is not rhetorical. GPU thermal design power is closing in on the 1,000-watt-per-chip threshold, a boundary that makes air cooling physically inadequate for high-density AI clusters. Power engineers are being asked to architect racks that may draw a full megawatt — a density that breaks the 54-volt DC power distribution standard before a single GPU runs a training job. And fiber operations teams are being asked to certify infrastructure at speeds and connector densities their test equipment may not have been qualified to handle six months ago.

Today's agenda is a real-time audit of where the industry's smartest practitioners actually disagree — and what the best available answers look like as of mid-2026.

Single-Phase vs Two-Phase: Why the Cooling Debate Is More Settled Than It Sounds

The session titled "1P vs. 2P: Inside the Liquid Cooling Divide Powering AI Data Centers" brings together Josh Claman of Accelsius, Kamal Mostafavi of CoolIT Systems, and Brian Kelly of Panduit. Framed as a debate, the practitioner math points in one direction more strongly than the panel's balanced billing suggests.

Two-phase direct-to-chip cooling works by circulating a dielectric or refrigerant fluid directly to a cold plate on the GPU surface. Rather than removing heat by warming a liquid, the fluid boils at the chip, exploiting the latent heat of vaporization to transfer heat with far less temperature rise. The physics are elegant two-phase DTC offers 10 to 100 times greater heat transfer capacity than the single-phase alternative. For chips approaching 1,000 watts, that capacity advantage is real.

Single-phase cold plate cooling — circulating a 75% water, 25% glycol solution to remove heat through convection — does not have that heat transfer ceiling advantage. What it has is everything else simpler deployment, lower upfront cost, easier integration with existing facility water systems, alignment with silicon roadmaps from every major GPU vendor, and, crucially, no PFAS exposure.

That last point now dominates the economics. Polyfluoroalkyl substances, the "forever chemicals" used in many two-phase dielectric fluids, face a TSCA Section 8(a)(7) reporting window set to open no later than January 31, 2027, under an EPA final rule issued April 13, 2026. State-level bans in New Jersey, Maine, Massachusetts, Minnesota, and Washington are already in place or planned. For a data center operator committing to infrastructure with a 10-to-15-year service life, that regulatory trajectory effectively prices in a fluid replacement cost that two-phase's heat transfer advantage does not recover.

The result single-phase DTC now commands an estimated 55% of the direct-to-chip market in 2026. AI datacenter liquid cooling as a whole is valued at approximately USD 3.7 billion in 2026, projected to reach USD 18.1 billion by 2036 at a compound annual growth rate of 17.2%.

The evaporative cooling session with Cody Weeks of Munters adds appropriate nuance: data halls, penthouses, office environments, gas turbine pre-cooling, and condenser pre-cooling all remain viable evaporative applications. The liquid cooling transition is real; it is not total.

Why Does Rack Power Beyond One Megawatt Make 54VDC Impossible?

The session "Powering the Megawatt Era: The Role of 800VDC in the Future of Data Center Energy," featuring Mike Tu of NVIDIA and Jim Simonelli of Schneider Electric, addresses what may be the most structurally significant infrastructure decision of the next five years.

The physics are straightforward. Power equals voltage times current (P = V × I). At 54 volts DC — the current in-rack distribution standard — delivering one megawatt of power requires approximately 18,500 amperes of current. That current must flow through copper busbars. At that scale, NVIDIA's own technical documentation notes that the busbars would consume up to 64 rack units of space for a Kyber-class megawatt rack, leaving no room for the compute hardware the rack is supposed to house.

At 800 volts DC, the same megawatt of power is delivered at roughly 1,250 amperes — approximately 15 times less current. The implications cascade: copper use is cut by an estimated 45%, the number of power conversion stages is reduced, end-to-end efficiency improves from approximately 83% to 92% or more, and total cost of ownership drops by roughly 30% per NVIDIA's projections. The analogy NVIDIA draws to the EV industry's migration from 400-volt to 800-volt platforms is precise: the physics are the same.

The transition requires new hardware throughout the power chain. Gallium nitride and silicon carbide wide-bandgap semiconductors replace traditional silicon in power supply units. Solid-state transformers at the facility perimeter replace conventional magnetic transformers. The ecosystem of hardware partners supporting 800VDC architectures had reached 29 named companies by mid-2026. Vertiv, Schneider Electric, Eaton, and Delta Power Group have announced commercial products expected in the second half of 2026, timed to the NVIDIA Kyber rack launch timeline. Foxconn's 40-megawatt Kaohsiung-1 data center in Taiwan is already operational on 800VDC, demonstrating the architecture at production scale.

Stanley Mlyniec of VoltServer presents a complementary approach: Class 4 fault-managed power, which transmits high-voltage DC in digitally verified "packets" with safety mechanisms built into the distribution system rather than relying on conventional circuit breakers. Where 800VDC directly addresses the current-reduction imperative for megawatt AI racks, fault-managed power addresses the safety engineering problem of distributing high voltages in dense, human-occupied facilities.

Day Two broadens the power conversation further. Greg Castle of Schneider Electric's session "Behind the Meter: The New Power Play in Data Center Site Strategy" maps the procurement landscape that has become the primary site-selection constraint for large-scale AI deployments: onsite generation, microgrids, firm natural gas, modular power plants, small modular reactors, fuel cells, and the regulatory pathways that determine where any of this can actually be permitted and built on a timeline that matches AI demand.

What Happens When 400G Infrastructure Cannot Support 800G Traffic?

The session "Testing High-Density Fiber Links: Best Practices for VSFF and 400G+ Networks," delivered by Jim Davis of Fluke Networks, addresses what the conference program aptly labels an "unsung hero" of AI infrastructure: the fiber validation challenge that sits between a data center's hardware investment and its ability to run AI workloads at the speeds the hardware is designed for.

The underlying issue is a fiber count discontinuity. AI training racks equipped with 16 H100-class GPUs can generate substantial east-west traffic — GPU-to-GPU communication across the back-end fabric — that scales with each successive GPU generation. At 400G, that traffic runs over eight multimode fibers (400GBASE-SR4). Moving to 800G requires twice as many: either two MPO-8/12 connectors per link or 16-fiber very small form factor (VSFF) connectors such as the SN-MT or MMC form factors.

That doubling has a direct operational consequence: existing 400G multimode infrastructure cannot support 800G links without recabling or connector replacement. In a data center that committed to its current fiber plant within the last two to three years, that represents a significant unplanned capital event.

The IEEE 802.3dj standard, which defines 1.6T Ethernet using 200 gigabits per lane (compared with the 100-gigabit-per-lane signaling in the existing IEEE 802.3df 800G standard approved in February 2024), is nearing ratification as of mid-2026. Broadcom and other manufacturers have been shipping pre-standard 1.6T transceivers since September 2024, meaning the standard's formal completion largely ratifies what deployment has already begun. The architecture difference matters: 802.3dj achieves 800G over four 200G lanes rather than eight 100G lanes, reducing pin counts, PCB routing complexity, and power per bit transmitted — but it currently covers only single-mode fiber, a consideration for data centers that made multimode infrastructure investments.

Nokia's head of Portfolio Marketing, Rob Shore, put the transition state plainly earlier this year 400G is now mainstream while 800G is ramping rapidly in 2026.

VSFF connectors — the SN-MT and MMC form factors — offer a path forward for operators who need 16-fiber density in the same physical footprint as current 8-fiber MPO connectors. But inspection and certification requirements differ. Angled physical contact (APC) fiber end-face standards, tighter cleanliness tolerances for single-mode APC applications in 800G DR links, and reflectance limits that the IEEE specifies to protect the low-cost transceivers in short-reach DR applications — each requires updated test procedures that many operations teams have not yet standardized.

The longer trajectory is already visible from the conference agenda: back-end inter-cluster connections are migrating from 800G toward 1.6T and will eventually target 3.2T, while front-end switch-to-switch and switch-to-server links are consolidating around 800G with 1.6T as the near-term target. Getting fiber validation right at 400G and 800G is foundational to every subsequent speed transition.

Prefabrication Is Not Just Faster — It Changes What Gets Built

Thomas Humphrey of Schneider Electric's session "The New Build Model for AI: How Prefabrication Is Changing Speed to Market" addresses one of the most acute operational constraints the industry faces the gap between when a hyperscaler or colocation operator commits capital to AI capacity and when that capacity can generate revenue.

The problem is structural. Traditional data center construction sequences site work, structural steel, mechanical, electrical, and IT infrastructure in phases that stretch 18 to 36 months from site approval to energized compute. AI demand curves do not accommodate 36-month build cycles. Modular and prefabricated approaches shift substantial assembly work to controlled factory environments, compressing integration timelines and reducing on-site commissioning risk. The approach has moved from experimental to mainstream among hyperscale operators and AI-focused colocation providers deploying at speed. Hyperscalers — Amazon, Microsoft, Google, Meta, and Oracle — are projected to spend more than USD 600 billion on infrastructure in 2026, with approximately 75% of that targeting AI.

The closing session with Tom Carroll of ebm-papst Americas — "Designing for the AI Era: Hybrid Cooling, HVDC Readiness, and Resilient Data Center Infrastructure" ties these threads together hybrid cooling integration, power efficiency under high-voltage DC architectures, system controls, supply chain risk, noise and community impact, and the overarching question of resilience in facilities where downtime consequences are measurably higher than in prior generations of compute infrastructure.

Why Every Infrastructure Choice Made This Year Will Be Paid for in Operating Costs by 2030

Steven Carlini of Schneider Electric opens Data Center Insights 2026 with "Expert Roundup: The State of the Data Center Industry" — a session framing that is deceptively ordinary for a moment that is anything but.

AI workloads have changed nearly every foundational data center design assumption simultaneously. The cooling choice (which fluid, which topology, which regulatory exposure) has a 10-to-15-year service life. The power architecture choice (54VDC that cannot serve megawatt racks, or 800VDC that requires a new power chain) locks in efficiency, operating cost, and expansion headroom for the facility's entire useful life. The fiber infrastructure choice (16-fiber VSFF for 800G density, or two-connector workarounds) determines whether a data center can serve 1.6T when that standard matures — without rewiring.

The practitioners attending today's sessions are not choosing between options that work equally well. They are choosing between options that will prove right or wrong when the bills for the wrong choices arrive between 2028 and 2032. Data Center Insights 2026's format — cross-functional, multi-brand, engineer-targeted — is one response to a moment when no single technology domain holds the answer to converging constraints.

Sessions run 11:00 AM to 3:30 PM ET today and tomorrow. Registration for the virtual event remains open at the Data Center Insights 2026 event portal.

Frequently Asked Questions

What is the difference between single-phase and two-phase direct-to-chip liquid cooling, and which should a data center operator choose in 2026?

Single-phase direct-to-chip cooling circulates a glycol-water mixture (typically 75% water and 25% glycol) through cold plates mounted directly on GPU and CPU chips, removing heat through convection as the fluid warms. The fluid stays liquid throughout the cycle. Two-phase DTC uses a dielectric or refrigerant fluid that boils at the chip surface — exploiting the latent heat of vaporization to transfer heat far more efficiently, at 10 to 100 times the capacity of single-phase systems. The problem with two-phase in 2026 is regulatory: the "forever chemicals" (PFAS) used in many two-phase dielectric fluids are subject to an EPA TSCA Section 8(a)(7) reporting window opening no later than January 31, 2027, with state-level bans already in force in several states. For most operators committing infrastructure with a 10-to-15-year service life, single-phase DTC is the lower-risk choice today — which is why it holds an estimated 55% of the direct-to-chip market in 2026.

Why is 800VDC considered necessary for AI data centers, and when will it be commercially available?

At the current 54VDC in-rack distribution standard, delivering one megawatt of rack power requires roughly 18,500 amperes of current. NVIDIA has documented that this current load would require busbars consuming up to 64 rack units of space at megawatt scale — leaving no room for the compute hardware. At 800VDC, the same power is delivered at roughly 1,250 amperes, cutting copper use by approximately 45% and improving end-to-end efficiency from around 83% to 92% or more. Commercial products from Vertiv, Schneider Electric, Eaton, and Delta are expected in the second half of 2026, aligned with the NVIDIA Kyber rack timeline.

What are VSFF connectors, and why does upgrading from 400G to 800G fiber require recabling?

VSFF (very small form factor) connectors — the SN-MT and MMC designs — provide 16-fiber connectivity in roughly the same physical footprint as a traditional 8-fiber MPO connector. The need for 16 fibers at 800G is a fundamental result of IEEE standards: the current 400GBASE-SR4 standard runs 400G over 8 fibers at 100 Gbps per lane. The 800G upgrade requires twice the fiber count. Existing 8-fiber multimode runs physically cannot support 800G without either using two connectors per link (a density penalty) or replacing connectors with VSFF designs. Data centers that wired

Source

• techtimes.com (2026-07-15)
• Original article: AI Rack Limits Exposed: Cooling, 800VDC, and Fiber at Data Center Insights 2026 Today