Weekly Industry Pulse: Liquid Cooling Becomes Default, Jetson Thor Replaces Orin & the Data-Center Power Crunch

1. Liquid Cooling: From Optional Upgrade to the Default Delivery Mode

This week, the industry stopped asking whether liquid cooling is coming and started treating it as the baseline assumption for any serious GPU deployment. The numbers have become impossible to argue with.

The Numbers

What Changed This Week

• Direct-to-chip is displacing indirect cooling. Cold plates and direct-to-chip designs are winning market share from indirect (in-row / in-rack) approaches — indirect's share falls from 32.5% in 2025 to roughly 24% by 2034. The fastest growth is in small and mid-sized data centers (33.9% CAGR), not hyperscalers — which means liquid cooling is now a real line item for enterprise and edge buyers, not just the cloud giants.
• Rack-level integration is the new unit of sale. Supermicro's 48U reference rack packs 64 current-generation GPUs across eight 4U 8-GPU servers plus a CDU, with HGX B200 as the workhorse. Customers increasingly quote "whole-rack liquid delivery," not bare GPU servers.
• Shared-memory models are driving the physics. CoreWeave frames the reason plainly: an 8-GPU HGX H200 node already presents 1.13 TB of shared memory, and models scaling 10× will need 10 TB+ — a memory-and-bandwidth wall that air cooling cannot get past.

Buyer Implications

If you are procuring B200-, H200-, or Rubin-class GPUs, budget liquid cooling as a default line item, not an optional add-on. Expect vendor RFPs to lead with cold-plate / CDU capability and rack-level thermal delivery. For edge and mid-size clusters, the 33.9% CAGR segment means CDUs and cold plates are now available at realistic MOQs and lead times — direct-to-chip is no longer hyperscaler-only. Buyers who still spec air-cooled high-density nodes are buying into a thermal ceiling that next-gen silicon will exceed within one refresh cycle.

2. Edge AI Hits the Mass-Deployment Inflection — and Jetson Thor Replaces Orin

The second theme of the week: edge AI is no longer a proof-of-concept market, and the platform transition from Jetson Orin to Jetson Thor is now underway in force.

The Market Data

Jetson Thor: The Generational Handoff Is Real

The clearest signal this week was the industrial rollout of Jetson Thor, NVIDIA's flagship edge platform after Orin:

• FORECR launched the DSBOX-THRMAX, a fanless industrial edge AI computer built on Jetson Thor with 128 GB LPDDR5X and QSFP+ networking — aimed squarely at robotics, machine vision, and predictive maintenance.
• ARBOR, Premio, and Syslogic all announced Jetson Thor industrial Box PCs, following the same "SoM + hardened chassis" pattern that made Orin the industrial standard.
• MSI at Embedded World 2026 showed the EdgeXpert AI Supercomputer (NVIDIA GB10 Superchip) alongside the MS-C926 industrial PC — evidence that system vendors are packaging edge AI as complete ecosystems, not bare modules.

Source: FORECR · Source: PR Newswire / MSI

The Buying Criterion Is Shifting — Away From Raw TOPS

The most important procurement insight of the week is not a product but a change in how buyers evaluate hardware. The Edge AI Vision Alliance's survey of 250 semiconductor executives makes it explicit:

• AI chips represent just 0.2% of chip unit production but roughly 50% of industry revenue — the value has migrated to the compute layer, and buyers are now scrutinizing it.
• The purchase metric is moving from peak FLOPS to per-watt efficiency, latency-per-query, and cost-per-inference. A small, well-optimized model on efficient hardware beats a big, power-hungry chip on TCO.
• Software — not silicon — is the bottleneck. NPU architectures are fragmented (AMD, Intel, Qualcomm, Apple all incompatible), so the buyers who de-risk fastest are the ones binding to a single software ecosystem (read: NVIDIA Jetson) rather than chasing bare NPU dies.

Source: Lattice Semiconductor (echoed by Wevolver's 2026 Edge AI Technology Report from Edge Impulse / MIPS / Murata / Synaptics / Synopsys)

Buyer Implications

Two takeaways for procurement teams:

1. Plan the Orin → Thor transition now. The industrial ecosystem is already shipping Thor. If you are designing new carrier boards or committing to a multi-year edge program, verify Thor compatibility to avoid stranding Orin investments — but note that Orin Industrial's 248 TOPS at 15–75 W with a 10-year availability commitment remains the right choice for harsh-environment production deployments today.
2. Sell — and buy — on efficiency, not TOPS. Both QSCompute's own messaging and your internal evaluation should lead with watts-per-inference and cost-per-query. Software-stack readiness (JetPack / Triton / ONNX Runtime support) is worth more than another 50 TOPS on a fragmented NPU.

3. The Data-Center Power Crunch: Efficiency Is Now a Hard Constraint

A quieter but equally consequential thread ran through the week: power is becoming the binding constraint on AI buildouts, and it is quietly rewiring how buyers evaluate every component in the stack.

What the Data Says

• Gas turbine orders are booked through 2028, and analysts warn data centers face a real risk of power unavailability before 2028 — new AI capacity is increasingly gated by grid access, not GPU supply.
• Rack power density is compounding the problem. At 27 kW and climbing (+69% YoY), a single modern rack now draws what an entire 2020-era data-center aisle did — which is precisely why liquid cooling moved from option to default this week.
• The hedge is efficiency, not just more megawatts. The same forces driving liquid cooling adoption — per-watt compute, direct-to-chip thermal removal, and edge offload — are the practical responses to a grid that cannot keep pace.

Source: Supermicro · Source: IEEE Spectrum — Data Center Liquid Cooling: The AI Heat Solution

Buyer Implications

Power constraints mean three things for procurement right now:

• Efficiency is a negotiation point, not a footnote. When a data center cannot get more power, the buyer who can deliver more compute per kilowatt wins the deal. Spec sheets should lead with performance-per-watt.
• Edge offload becomes a capacity strategy. Deploying inference at the edge (the 20% penetration inflection above) is not just a latency play — it is a way to buy compute capacity that does not depend on a constrained data-center power contract.
• Lock in thermal and power infrastructure early. Liquid cooling components (cold plates, CDUs, quick disconnects) and high-efficiency power supplies are on the same 2028-critical path as GPUs. Treat them as long-lead items, not accessories.

What We Learned