OCP Global Summit 2026 | Arista | Andy Bechtolsheim | The Gigawatt AI Data Center Era

1. Why Arista's Bechtolsheim Is Talking About Buildings

Bechtolsheim is chief architect of a networking company, yet this talk barely touched switches. It was about concrete, transformers and drones. That alone is a signal: for AI data centers, the bottleneck has moved up from chips and networks to buildings, power and construction time.

He opened with scale (BloombergNEF projection): about 250 GW of new data center capacity worldwide from 2026 to 2035, with 150 GW in the US, 85 GW in China and 15 GW in Europe. He added verbally that what is being planned is roughly twice that.

Here is how today's gigawatt campuses are built:

Data center	Compute power	Building size (sq ft)	Density (kW/sq ft)
xAI Colossus	400 MW	785,000	0.50
Stargate (Phase 1)	1.2 GW	4,000,000	0.30
Hyperion (Phase 1)	2 GW	4,000,000	0.50
Hyperion (Phase 2)	5 GW	10,000,000	0.50
Next-gen data center	5 GW	1,000,000	5.00

These sites were optimized for 120 kW liquid-cooled NVL72 racks. Bechtolsheim's argument is blunt: racks are heading to 600–800 kW, while buildings are still designed around 120 kW.

2. Density Breakdown: From a 2.3 MW POD to a 28.8 MW POD

Today's standard building block is a 2.3 MW POD: 16 GPU racks at 127 kW plus 16 network and storage racks at 14 kW.

Bechtolsheim's next-gen 28.8 MW POD occupies 2,772 sq ft (42 by 66 ft), with compute-area density above 10 kW/sq ft:

• 32 GPU racks × 800 kW = 25.6 MW (144–288 GPUs per rack)

• 16 networking racks × 125 kW = 2.0 MW

• 16 storage racks × 62.5 kW = 1.0 MW

• 16 CDUs × 12.5 kW = 0.2 MW

▲ Power split of a 28.8 MW POD: GPU racks take 89%, network racks 6.9%, storage 3.5% and CDUs 0.7% (Image source: Simple Tech Trend; data from Arista's 2026 OCP Global Summit presentation)

Add the transformers, batteries and breakers, and a "power plus compute" POD takes 5,544 sq ft at 5 kW/sq ft. Eight of them make 230.4 MW (44,352 sq ft, 67 m maximum fiber, 670 ns round trip); 32 of them make 921.6 MW in 177,408 sq ft, roughly 4 acres rather than thousands.

▲ Today's gigawatt-class data centers run at 0.3–0.5 kW per square foot; Arista's next-gen POD reaches 5 kW per square foot, so 921.6 MW needs only 177,408 square feet (Image source: Simple Tech Trend; data from Arista's 2026 OCP Global Summit presentation)

3. Where the Network Sits: 7% of the Power, One-Tenth of the Latency

For optics readers, the talk carried plenty of network numbers; they were just hidden in the floor plans.

Switch rack density jumps 8x. The slide puts today's air-cooled switch rack at 30 kW and 800 Tbps, and the next-gen liquid-cooled rack at 125–250 kW and 6.5 Pbps. Bechtolsheim said verbally this takes "liquid cooling and XPO modules" (ASR transcript, to be verified). We covered XPO's positioning in OCP APAC Summit 2026: the CPO / NPO / XPO panel: it is Arista's bet on high-density pluggables, not CPO.

▲ Switch racks move from 30 kW, 800 Tbps air-cooled to 125–250 kW, 6.5 Pbps liquid-cooled (Image source: Arista | 2026 OCP Global Summit presentation (02:50))

One number worth unpacking (STT calculation): 30 kW over 800 Tbps is about 37.5 W/Tbps; 125 kW over 6.5 Pbps is about 19 W/Tbps, and at 250 kW it returns to about 38 W/Tbps. Rack bandwidth density rises mainly because liquid cooling lets more power in; energy per bit does not fall in proportion. That is why switch thermal design is becoming its own topic, see 51.2T NPO switch thermal and cooling.

Latency is the most direct dividend density pays the network. The slide's rule of thumb is about 1 µs round trip per 100 m of fiber; the 921.6 MW site tops out at 220 m, or 2.16 µs round trip. A footprint 10x larger means 10x the latency. Bechtolsheim stressed that this matters for both scale-up and scale-out.

But keep one distinction: density cuts fiber length, not fiber count. The slide lists "1/10th the length of fiber"; with the same GPU count, the topology needs just as many fibers. Squeezing the same fiber count into one-tenth the area makes cabling denser, not lighter. As we argued in What killed slow-and-wide was 60,000 fibers, not physics, fiber count and connectivity are a challenge in their own right.

4. Cost and Schedule Breakdown: $4B Saved per GW

Bechtolsheim called these "not scientific numbers," but the order of magnitude is clear:

Construction item	Conventional	Labor share	Power dense
Electrical infrastructure	$5B	30–40%	$3.6B
Cooling infrastructure	$3B	35–45%	$2B
Building shell, civil	$1.5B	40–50%	$0.2B
Commissioning, management	$0.5B	65–80%	$0.2B
Total per GW	$10B		$6B

▲ Data center shell cost per GW: $10 billion conventional versus $6 billion high-density, excluding GPUs and networking (Image source: Arista | 2026 OCP Global Summit presentation (07:10))

The most telling part of the table is where the savings land: the building shell drops from $1.5B to $0.2B, nearly 90% off, while electrical only falls from $5B to $3.6B. In other words, density saves concrete and labor; most of the money for power equipment still has to be spent.

Schedule matters even more. The slide says a gigawatt AI data center takes at least 24 months from breaking ground to going live (excluding permits), and on-site labor for shell, electrical and mechanical work is now the critical path; a 10x denser design could halve that. With GPUs on a yearly cadence, waiting 12 fewer months is worth more than saving $4B.

5. Power: The Precondition for Density

Without new power components, 10x density is just a floor plan. Bechtolsheim's building blocks:

• 9.6 MW solid-state transformer (SST): 34.5 kV medium voltage in, 800 VDC out, 98.7% efficient, liquid-cooled, about half a shipping container (slide example: Luxtech Nova Link). Each POD uses four for N+1.

• 6 MWh battery: CATL Tener-S LFP in a 20 ft container, 0.5P discharge, 3 MW for 2 hours. His argument is that most outages last under 2 hours, which could remove on-site generators and speed up permits.

• 2 MW CDU: CoolIT CHx2000 in a single-rack footprint, drawing 12.5 kW.

• Solid-state circuit breaker (SSCB): conventional breakers take tens of milliseconds and DC has no zero crossing, making arc flash the most dangerous event in a data center; SSCBs use silicon carbide to clear a fault within 10 µs. Each 600 kW rack is fed by four 800 V circuits, each protected by a 250 A SSCB (200 kW); higher-power racks will need 400 A.

The 800 VDC efficiency gap is also clear: end-to-end efficiency to 12 V is 0.96 versus 0.93 for 480 VAC, about 3% saved, or 30 MW per GW.

6. Resilience: Data Centers as Fortresses

In the second half, Bechtolsheim took on an unusual topic: drone attacks. The slides cite attacks on multiple UAE data centers in March 2026 and split the threat in two: 5–100 kg high explosives (blast overpressure and fragments) and 2–4 kg shaped charges (able to penetrate 12 in of steel or more than 48 in of concrete). Damage could take a facility offline for 6 months to over a year.

His answer is a five-layer structure: steel mesh net (50 ft from the wall), steel slat armor (50 ft), an outer ultra-high-performance concrete (UHPC) wall, a 30 ft filler gap and an inner UHPC wall. For materials he favors carbon-nano-fiber concrete (CNFC): $960 per cubic meter versus $120, but with half the volume and lower rebar and placement costs, total placed cost is only 15% higher and installation time is 0.5–0.8x; a 24 in slab can shrink to 12 in.

▲ Five-layer protection design: steel mesh, steel-bar armor, outer UHPC wall, fill layer and inner UHPC wall, aiming for zero penetration into the data hall (Image source: Arista | 2026 OCP Global Summit presentation (16:14))

This section and the density argument are the same point: a smaller building is easier to harden.

7. Risks and Counterarguments

• The numbers are estimates, not quotes. Bechtolsheim himself called the shell cost table and the $4B per GW saving rough estimates, and the labor-share ranges are wide.

• Key components are not in volume yet. He described the 9.6 MW SST, high-current SSCBs and CNFC as in development or "still in the lab," so the density timeline depends on their shipments.

• 800 kW racks are themselves a precondition. The POD's 25.6 MW assumes 800 kW per rack; if rack power stalls at 600 kW, density has to be discounted.

• Area is not the only limit. Grid capacity, permits and water are the real ceilings for many campuses, and a smaller building does not make gigawatt power easier to secure.

• Fiber cabling gets denser. With the same fiber count in a smaller area, cabling, cleaning and maintenance challenges concentrate, and the talk offered no solution.

8. What It Means for Taiwan's Supply Chain

• None of the suppliers named in the talk is Taiwanese. The SST example is Luxtech, the battery is CATL, the CDU is CoolIT and the SSCB example is Atom Switch. Bechtolsheim's reference design does not currently use Taiwanese power vendors as its template.

• 800 VDC is where Taiwan's power and thermal vendors are competing at the same event. At this OCP Global Summit, Delta (800 VDC power and cooling), Lite-On (800V-to-chip) and Auras (liquid cooling) all gave dedicated talks; whether they can move into SSTs, SSCBs and 800 V rack power, the components density requires, is the direction to watch over the next 12–24 months.

• Liquid-cooled switch racks are a new spec on the network side. A 125–250 kW switch rack means switches and optical modules both need cold plates, a new design requirement for Taiwanese thermal vendors and switch ODMs, but the talk named no suppliers, so for now it is a direction to track.

9. Conclusion

Bechtolsheim's talk reorders the AI data center problem: power and buildings first, chips and networks second. His answer is 10x density: 921.6 MW in 4 acres, 220 m maximum fiber, $4B saved per GW and potentially half the build time.

For networking and optics, the design cuts both ways. The good news is that latency and fiber length both drop to one-tenth; the bad news is that fiber count stays the same, racks must be liquid-cooled and energy per bit does not fall in step. Over the next 12–24 months, watch three things: when 9.6 MW SSTs and high-current SSCBs reach volume, when 800 kW racks ship, and who breaks ground on the first campus built at this density.

Once racks reach 800 kW, the data center race is no longer about whose chip is faster, but about who can build 1 GW in a year.

This article is for technology and industry trend analysis only and does not constitute investment advice.