Cerebras Blocks Trump’s TPU Deal? No — Nvidia’s NVL144 in 2026

Nvidia’s stock has been pinned around the $200 mark through an entire Big Tech earnings cycle, and the hot takes blame a demand cliff — Cerebras winning inference share, Google’s TPU deal reshuffling the deck, some vague “AI capex digestion” narrative. That’s not what’s happening. The real constraint is physical: Nvidia Vera Rubin NVL144 power requirements land in the 200-250kW-per-rack range, and the overwhelming majority of datacenter shells coming online in 2026 were engineered for 30-40kW per rack. You cannot ship a rack into a building that can’t feed it or cool it. That retrofit math — not order books — is what’s quietly capping the 2026 shipment curve.

What’s actually new about Nvidia Vera Rubin NVL144 power requirements

Vera Rubin succeeds Blackwell Ultra, pairing the Rubin GPU with Nvidia’s own Vera CPU, an Arm-based custom core replacing Grace. The NVL144 designation means 144 GPU dies in a single NVLink domain. Here’s the naming trap that snares half the analyst notes: NVL144 and GB300 NVL72 both contain 72 packages. Nvidia switched from counting packages to counting dies. NVL144 does not double the rack in socket count. It’s the same 72-slot compute tray topology with dual-die packages, a new NVLink 6 fabric at roughly 3.6TB/s per GPU, and HBM4 replacing HBM3E.

What did roughly double is the power envelope. GB300 NVL72 racks draw about 120-140kW. Vera Rubin NVL144 is guided toward 200-250kW depending on configuration, with the Rubin Ultra NVL576 generation in 2027 pointed at 500kW+. That’s the number that breaks things. A 250kW rack is not a bigger version of a 40kW rack — it’s a different electrical class. It requires 800VDC power distribution instead of traditional 415VAC-to-rack-PDU topologies, because at 250kW the copper busbar mass and I²R losses in an AC scheme become absurd. Nvidia publishes 800VDC reference architectures with partners like Vertiv, Delta, and Schneider precisely because the ecosystem doesn’t exist yet.

Cooling is the second wall. At 250kW per rack, air does nothing useful. Direct-to-chip liquid is mandatory, and the coolant distribution unit (CDU) capacity behind those racks becomes the binding constraint. A typical row-level CDU in 2026 handles 1-2MW. Feeding eight NVL144 racks needs 2MW of CDU capacity, plus facility water loop, plus a cooling tower or dry cooler sized for it, plus structural floor loading for racks that weigh close to two tons. Most 2026 shells — leases signed in 2023 and 2024, designed against Hopper-era assumptions — have none of this. The retrofit is a 12-18 month capital project, not a firmware update.

Why it matters

• Shipment guidance is gated by buildings, not fabs. TSMC CoWoS capacity and HBM4 supply are real constraints, but money solves them. Rewiring a live datacenter to 800VDC takes money and 18 months, and the second variable doesn’t compress.
• The “Cerebras/TPU is stealing share” narrative measures the wrong thing. Alternative silicon isn’t winning on merit so much as on deployability — a wafer-scale or TPU pod that fits an existing power envelope ships today. That’s a temporary structural advantage, not an architectural one.
• NVL144 rack density forces a split fleet. Operators will run GB300 NVL72 in retrofitted-lite halls and reserve Vera Rubin for greenfield 800VDC builds. Expect two-tier pricing and two-tier availability for at least six quarters.
• Rubin CPX changes the unit economics of inference. Nvidia’s CPX variant strips HBM in favor of cheaper GDDR7 for the prefill/context phase, targeting long-context inference at far better perf-per-dollar. A Rubin CPX inference rack carries a materially different power profile than a training NVL144, and it’s the SKU most likely to land in constrained facilities.
• Power procurement becomes a competitive moat. Whoever locked substation interconnect queue positions in 2023-2024 gets to deploy in 2026. Grid interconnect backlogs in Northern Virginia, Dublin, and Santa Clara run 4-7 years. That’s a real barrier to entry.
• Your model deployment plan needs a facility line item. If you buy capacity rather than build it, “when can you actually rack this” is now a harder question than “what’s the hourly rate.”

How to use it today: auditing your own power and thermal headroom

Whether you operate a colo cage or rent capacity, you can quantify the gap before signing anything. Here’s a practical sequence.

1. Establish your actual per-rack ceiling, not the marketing number. Pull real draw from your PDUs rather than trusting the design spec — most halls derate in practice.

   # Query a rack PDU over SNMP for real-time load (Raritan/ServerTech MIBs vary)
   snmpwalk -v2c -c public 10.20.30.41 \
     1.3.6.1.4.1.13742.6.5.4.3.1.4 | tail -20
   
   # Redfish equivalent — modern PDUs and BMCs both expose this
   curl -sk -u admin:$PDU_PASS \
     https://10.20.30.41/redfish/v1/PowerEquipment/RackPDUs/1/Outlets \
     | jq '[.Members[].PowerWatts] | add'
   

2. Model the retrofit gap in explicit terms. This arithmetic decides whether Vera Rubin is even an option for you.

   # nvl144_gap.py — how many racks can this hall actually take?
   HALL_UTILITY_KW      = 4000    # committed utility feed to the hall
   PUE                  = 1.15    # liquid-cooled target
   EXISTING_LOAD_KW     = 2600    # what's already racked
   NVL144_RACK_KW       = 250     # Vera Rubin NVL144, mid config
   GB300_RACK_KW        = 132     # GB300 NVL72 for comparison
   CDU_CAPACITY_KW      = 2000    # installed liquid loop capacity
   
   it_headroom = (HALL_UTILITY_KW / PUE) - EXISTING_LOAD_KW
   power_limited  = int(it_headroom // NVL144_RACK_KW)
   cooling_limited = int((CDU_CAPACITY_KW - EXISTING_LOAD_KW * 0.9) // NVL144_RACK_KW)
   
   print(f"IT headroom:        {it_headroom:.0f} kW")
   print(f"Power-limited NVL144 racks:   {max(power_limited, 0)}")
   print(f"Cooling-limited NVL144 racks: {max(cooling_limited, 0)}")
   print(f"Actual deployable:            {max(min(power_limited, cooling_limited), 0)}")
   print(f"Same headroom in GB300 NVL72: {int(it_headroom // GB300_RACK_KW)}")
   

   Run that against your own numbers. Cooling usually binds before power, and the same headroom absorbs roughly twice as many GB300 racks — which is exactly why NVL72 demand isn’t collapsing.

3. Check thermal margin on what you already run. If your current fleet already throttles, you have no headroom story to tell.

   nvidia-smi --query-gpu=index,name,power.draw,power.limit,temperature.gpu,clocks_throttle_reasons.active \
     --format=csv,noheader -l 5
   
   # Sustained hardware slowdown across the fleet = cooling is already the limiter
   dcgmi dmon -e 1001,1004,1005,203,155 -c 20
   

4. Cap power to fit the envelope you have. A blunt but effective bridge — you trade throughput for the ability to rack at all.

   # Trim each GPU's board power to fit a constrained rack budget
   sudo nvidia-smi -pm 1
   sudo nvidia-smi -pl 1000          # watts per GPU; adjust to your rack ceiling
   
   # Kubernetes: keep scheduling honest about the reduced envelope
   kubectl label node gpu-node-07 nvidia.com/power-profile=capped-1000w
   

5. Put the facility question in your procurement prompt. When evaluating colo or cloud contracts, demand these specifics in writing:

   Evaluate this datacenter capacity proposal against Vera Rubin NVL144 deployment.
   
   Extract and flag as MISSING if absent:
   - Committed per-rack power ceiling (kW) and whether it is N or N+1
   - Power distribution topology: 415VAC PDU vs 800VDC busbar
   - Installed CDU capacity (kW) and secondary loop delta-T
   - Rack floor loading limit (kg/rack) vs ~1,800kg for NVL144
   - Utility interconnect status and any queue position date
   - Contractual date liquid-cooled capacity is energized, not "available"
   
   Output a table: requirement | stated value | gap vs NVL144 | risk (H/M/L).
   

How it compares

Platform	Rack power (approx.)	Cooling	Power topology	2026 deployability
Nvidia GB300 NVL72	120-140kW	Direct-to-chip liquid	415VAC / 48VDC rack bus	High — fits retrofitted halls
Nvidia Vera Rubin NVL144	200-250kW	Liquid, higher CDU demand	800VDC preferred	Low — greenfield or deep retrofit
Nvidia Rubin CPX (inference)	Lower per-rack; GDDR7, no HBM	Liquid, reduced load	Flexible	Medium-high — the constrained-facility SKU
Google TPU v7 pod	Facility-integrated, not rack-sold	Google’s own liquid loop	Purpose-built halls	High — Google controls the whole stack
Cerebras WSE cluster	~23kW per CS-3 chassis	Internal closed loop	Standard AC	High — drops into conventional space
AMD MI450 / Helios rack	~150-190kW (announced)	Liquid	Moving to 800VDC	Medium — same wall, slightly lower

Read that table as a deployability ranking, not a performance one. On raw training throughput per rack, Vera Rubin wins decisively. On racks you can actually energize in calendar 2026, it sits near the bottom — and revenue recognition follows energization, not benchmarks.

What’s next

Watch three signals through 2026. First, 800VDC ecosystem maturity: when Vertiv, Delta, Schneider, and Eaton all ship volume 800VDC busway and rectifier product with lead times under 20 weeks, the retrofit bottleneck starts to clear. Right now, transformer and switchgear lead times stretch past 100 weeks in some markets, and that number gates the buildout. CDU supply compounds it — liquid cooling CDU capacity in 2026 is a thin supply chain with maybe a half-dozen credible vendors at scale.

Second, watch how aggressively Nvidia pushes Rubin CPX. If the company leans into the inference SKU with lower thermal demands, that’s a tacit acknowledgment that the flagship NVL144 can’t ship into the installed base fast enough. CPX is the pressure-release valve: it monetizes long-context inference in facilities that will never host a 250kW rack. A strong CPX ramp is bullish for revenue and simultaneously confirms the density thesis.

Third, the Rubin Ultra NVL576 at 500kW+ in 2027 means the industry doesn’t get to catch its breath. Every facility designed today to handle 250kW is already a generation behind what lands 18 months later. The rational operator designs for 600kW-capable infrastructure now and eats the overbuild cost — which is exactly why hyperscaler capex guidance keeps climbing even as people insist we’re near a top. That capex isn’t going into GPUs alone; a growing fraction is buildings, substations, and water.

Frequently Asked Questions

Does NVL144 mean 144 GPUs per rack, double the NVL72?

No, and this is the single most common error in coverage. NVL144 counts 144 GPU dies across 72 dual-die packages, in the same 72-slot rack topology as GB300 NVL72. Nvidia changed the counting convention. The compute uplift is real, but it comes from a new architecture, HBM4, and NVLink 6, not from stuffing twice the sockets into a rack.

Why is 800VDC necessary instead of the existing AC distribution?

At 250kW per rack, delivering power at conventional voltages means enormous conductor cross-sections and significant I²R losses in the busbar, plus multiple AC-DC conversion stages each shedding efficiency. Moving to 800VDC at the row or hall level cuts conductor mass roughly proportionally, eliminates conversion stages, and improves end-to-end efficiency by several points. At 40kW per rack none of this mattered; at 250kW it dominates the design.

Can I retrofit an existing 40kW-per-rack hall for Vera Rubin?

Technically yes, practically it’s close to a rebuild. You need new power distribution, a liquid loop with CDUs and secondary piping, heat rejection sized for the new load, possibly structural floor reinforcement, and — the hard part — a bigger utility feed. That last item depends on an interconnect queue you don’t control. Budget 12-18 months minimum, and confirm the utility timeline before anything else.

Is Cerebras or Google’s TPU actually taking share from Nvidia?

They’re taking deployments Nvidia physically can’t fill in the window, which looks identical to share loss on a quarterly chart. Cerebras drops into standard facilities; Google’s TPU pods live in halls Google designed around them. Neither wins a like-for-like performance comparison against Rubin. The advantage is scheduling, and it persists only as long as the datacenter power constraint does.

What is Rubin CPX and why does it matter for constrained facilities?

Rubin CPX is a Rubin variant built for the prefill/context phase of inference. It swaps expensive HBM for GDDR7, cutting cost and power substantially while remaining well-suited to compute-bound long-context prefill. For an operator who cannot energize a 250kW training rack, a Rubin CPX inference rack is often the only Rubin-generation hardware they can realistically deploy — and it targets one of the fastest-growing workload classes.

If demand is fine, why has the stock been flat?

Near-term revenue is a function of racks energized, not orders booked, and the market can’t easily model a constraint that lives in utility interconnect queues and CDU vendor lead times. Backlog keeps growing while quarterly recognition stays gated by facilities. That’s a timing story that looks like a demand story on a price chart, and it resolves as 800VDC-ready capacity comes online, not as a function of anything Cerebras or Google announces.