Data Center Market Update: Week of August 17, 2026

This week’s news is regulatory and structural rather than a single mega-deal: FERC’s 17 August deadline forced all six US grid operators to show their large-load interconnection rules are fit for purpose, Dell’Oro put global data center capex above $3 trillion by 2030, and Texas set a firm 10 December date for its ERCOT audit of pending data center interconnections. Outside the US, Johor added another gigawatt-scale campus and a long-haul fiber deal underscored how much AI buildout now depends on network capacity, not just power and racks.

Key takeaways

• All six FERC-jurisdictional grid operators (PJM, MISO, SPP, CAISO, ISO-NE, NYISO) faced a 17 August 2026 deadline to justify or reform large-load interconnection tariffs under FERC’s June show-cause orders — PJM and SPP already had reforms in motion, MISO and NYISO are further behind.
• Dell’Oro Group forecast on 18 August that global data center capex will exceed $3 trillion by 2030, with AI accelerators near a third of that spend and the top four US hyperscalers close to half of global capex.
• ERCOT told the Texas PUC on 20 August it will complete Governor Abbott’s data center interconnection audit by 10 December 2026, covering roughly 250-300 large loads already in the queue.
• YTL Power and JLand announced a 1.3-1.5GW Johor campus on 20 August, which combined with YTL’s existing Kulai park could reach 2.7GW; an IPO of YTL’s data center unit is under exploration for 2027.
• Zayo locked in long-term fiber supply from Corning for a 20-million-fiber-mile buildout through 2030, with 24 of 28 new routes tied to Nvidia-linked long-haul AI network capacity.
• H100 and B200 GPU rental prices held roughly flat week-on-week: H100 $1.49-6.98/GPU-hour, B200 $6.99-8.64/GPU-hour on marketplaces.

FERC’s large-load deadline: a real test of grid-operator reform speed

On 18 June 2026, FERC issued Section 206 show-cause orders to all six of its jurisdictional grid operators, preliminarily finding that existing interconnection tariffs may be unjust and unreasonable because they lack clear provisions for integrating data centers and other large loads above roughly 50MW. Each operator faced a 17 August deadline to either demonstrate its current rules are already adequate or file specific reforms. PJM and SPP entered the deadline pointing to reforms already accepted or in effect, while MISO and NYISO are working through longer internal stakeholder processes — NYISO’s planning reforms aren’t expected to reach a formal filing until around March 2027. For buyers, the practical effect is a widening gap in how fast new capacity can get grid access depending on region: see our grid connection queues guide for how queue position already varies by market.

Dell’Oro: data center capex crosses $3 trillion by 2030

Dell’Oro Group’s 18 August forecast puts cumulative global data center capital expenditure above $3 trillion by 2030, roughly in line with earlier 2026 estimates from JLL and Moody’s. AI accelerators account for close to a third of that figure, and the top four US hyperscalers — Amazon, Google, Meta and Microsoft — represent nearly half of global spend on their own. The scale reinforces why power and land, not construction labor, are now the binding constraints on new supply; see our data center investment guide for how that capital flows into wholesale and retail colocation pricing.

Texas sets a hard date on its ERCOT audit

ERCOT confirmed to the Texas Public Utility Commission on 20 August that it intends to file its comprehensive interconnection audit — ordered by Governor Abbott on 3 August — by 10 December 2026. The audit verifies information submitted by an estimated 250-300 large loads in ERCOT’s first interconnection batch (“Batch Zero”) and collects community-impact data from projects requesting 25-75MW. Any project that fails to comply with PUCT and ERCOT requirements risks losing its grid connection outright. Buyers with pending Texas capacity should treat the December filing date as a hard checkpoint on timeline risk; our due diligence checklist covers how to weight interconnection-status risk in a site evaluation.

Johor adds another gigawatt-scale campus

YTL Power International and JLand Group announced a partnership on 20 August to develop a second Johor campus with 1.3-1.5GW of planned capacity, which would sit alongside YTL’s existing Green Data Center Park in Kulai (up to 1.2GW). Combined, the two campuses could host up to 2.7GW of capacity in a single Malaysian corridor. YTL is also exploring a 2027 IPO of its data center unit that could value the business at roughly RM28 billion ($6.6 billion). The announcement adds to a pattern of Johor capacity growth documented in our Johor data center guide — the market now carries the largest incoming pipeline in Asia-Pacific alongside the region’s lowest colocation vacancy rate.

Location	Announcement (Aug 2026)	Capacity
Johor (YTL/JLand)	New partnership, 20 Aug	1.3-1.5GW
Johor (YTL Kulai, existing)	Green Data Center Park	Up to 1.2GW
Combined YTL Johor footprint	—	Up to 2.7GW

Fiber, not just power, is becoming a capacity bottleneck

Zayo announced a long-term fiber supply agreement with Corning around 20-22 August to secure capacity for a buildout of 28 new or overbuilt long-haul routes, roughly 20 million fiber miles, through 2030. Twenty-four of those routes are tied directly to Zayo’s work with Nvidia on long-haul network capacity connecting distributed AI training and inference sites. The deal is a reminder that AI infrastructure growth increasingly depends on middle-mile and long-haul network buildout, not only on-site power and rack capacity — a constraint that gets less attention than interconnection queues but affects how quickly distributed GPU capacity can actually be used as a single pool. See our cross-connect pricing guide for how network costs factor into total colocation spend.

GPU rental prices: little change week-on-week

GPU	Rate range (week of 17 Aug 2026)	Notes
H100	$1.49-1.99/GPU-hour (Vast.ai, RunPod) up to $6.98 (hyperscaler)	Roughly flat versus the prior week
B200	$6.99-8.64/GPU-hour (Nebius, Lambda, RunPod)	Blackwell access still gated by enterprise allocation at several clouds

With no major new GPU generation launched this week, pricing held within the same bands reported in our H100 rental price tracker and GPU cloud comparison.

What this means for buyers

Regulatory timelines are now as important as power availability when scoping a site: the FERC show-cause deadline and the ERCOT audit’s 10 December filing date both create concrete checkpoints where interconnection terms could change mid-negotiation, so buyers should build contract flexibility around both dates rather than treating current tariffs as fixed. In Southeast Asia, Johor’s capacity keeps compounding faster than most alternative corridors — worth benchmarking against Singapore and Bangkok before committing to a market. On networking, factor long-haul fiber lead times into any multi-site GPU deployment plan, not just local cross-connect costs. Cross-check current wholesale and GPU pricing against the index and get a live benchmark through our quote tool; browse underlying facility data in the data center catalog and track aggregate pricing trends in stats.