Waaree doubles Texas module capacity to 3.2 GW as cloud and memory demand squeeze supply

The Texas lines are not a pilot. Each carries 800 MW of annual capacity. Waaree says its board approved the expansion in April 2025. Output now feeds US developers, utilities and EPC contractors. The company frames the investment against Section 232 trade measures that have changed the economics of imported modules.

The financials behind it are worth reading closely. Revenue from Waaree's US manufacturing operations reached INR 13 billion ($146 million) in the second quarter of fiscal 2027, up from INR 7.32 billion a year earlier, according to pv magazine. The US domestic business accounted for 43.8% of the company's INR 592 billion order book, against roughly 36% at the end of the previous quarter. Consolidated revenue from operations rose 23.9% to INR 75.18 billion in the same quarter.

"Shifting trade policy has strengthened the case for local manufacturing, and that is exactly where we have invested," said Sunil Rathi, a director of Waaree Group, in comments carried by pv magazine. "With Brookshire now at 3.2 GW and Arizona expanding, we are serving American customers from American factories, and our US order book continues to grow every quarter."

Waaree is also expanding its Arizona module facility from 1 GW to 1.6 GW. The company expects that to take total US module capacity to 4.8 GW in fiscal 2027. Its global module order book exceeds 25 GW. In India, it plans to lift solar cell capacity from 5.4 GW to 15.4 GW in fiscal 2027. It also plans to commission a 10 GW ingot and wafer plant in fiscal 2028.

The other end of the chain: China moves offshore

On the same day, pv magazine reported that Chinese cell maker Yingfa Ruineng has unveiled a 7.5 GW plant in Indonesia producing n-type TOPCon cells. It is the company's first overseas manufacturing base, adding to three in China. The four together have a combined effective capacity of 51.5 GW, including 45.5 GW of n-type TOPCon, according to figures the company released.

The facility is pitched at Indonesia's 100 GW solar programme and at local content rules. "For project developers, contractors and module manufacturers, local cell production can offer shorter supply routes, closer technical coordination and greater flexibility in supply planning," the company said in a statement cited by pv magazine. It added that the local capability is intended to help Indonesian module makers meet the country's TKDN requirement, subject to certification.

Yingfa Ruineng ranked second for global solar cell shipments in the first half of 2026, according to the research firm InfoLink, behind only Tongwei. Two announcements on one day describe the same reorganisation of where panels and cells get made. One builds capacity inside the US tariff wall, the other outside China.

Compute is the tighter constraint

Solar capacity is expanding on published schedules. Data centre capacity is not. Worldwide spending on cloud infrastructure services reached $168 billion in the third quarter of 2026, up 55% year on year, according to a Synergy Research Group report covered by Data Center Knowledge. The market has doubled in eight quarters. AI-specific services now account for roughly one third of total revenue.

John Dinsdale, chief analyst at Synergy, told Data Center Knowledge that three factors drive the surge: capacity coming online, higher pricing, and a shift toward more expensive AI services. "A lot of the additional capacity that is coming online is specifically targeting more valuable, higher-priced AI services," he said. Amazon held 26% of worldwide cloud infrastructure revenue, Microsoft 19% and Google 16%. Microsoft and Google grew faster than Amazon. Neoclouds such as CoreWeave, Crusoe and Nebius were among the fastest growers outside the top three.

The construction pipeline behind that revenue is large but slow. CBRE's August report on North American data centre trends, cited by Data Center Knowledge, found 7,481 MW under construction across eight primary markets in the first half of 2026, up 24.8% from a year earlier. More than 80% was preleased, against 74.3% a year earlier. Power constraints kept pushing completion schedules out. JLL's August midyear report put hyperscalers at 59% of North American tenant demand, with 28 GW of owner-occupied capacity under construction, neoclouds at 11% and AI companies at 7%. Dell'Oro Group projected in August that global data centre capital spending would exceed $3 trillion by 2030, according to the same Data Center Knowledge piece. Those are forecasts, not orders.

What AWS says, and what analysts push back on

Amazon CEO Andy Jassy said AWS's large committed capacity is largely sold out through 2027, with a good chunk of 2028 already allocated. He said the company keeps headroom for On Demand, EC2 and Bedrock. InfoQ reports that Jassy made the remarks in a video podcast with Andreessen Horowitz co-founder Ben Horowitz and general partner Erik Torenberg, reported by Reuters on 23 October. "2027 is largely sold out," Jassy said, according to InfoQ. "We always leave a certain amount for on demand and our inference service bedrock, or EC2, but largely sold out all the big chunks. A good chunk of 2028 is sold out."

Cloud analyst David Linthicum disputed the framing. He told InfoQ that the reporting overstated the position. He drew a distinction between pre-reserved capacity blocks and what an enterprise can allocate day to day. "Of course, I suspect I can still go into the AWS console today and allocate GPU capacity for AI workloads," he said. "Maybe not the exact massive, pre-reserved, hyperscale capacity block some large model provider wants for the next three years, but that's not exactly the same thing as 'AWS is sold out.'"

"We've been seeing a big rise in compute capacity issues the past month or so at work, just for regular ec2 instances. Forget spot instances we have trouble getting enough on demand."

That comment, from a developer posting on Reddit as cbigsby and quoted by InfoQ, points at ordinary instance types rather than frontier training clusters. InfoQ also noted that AWS raised Capacity Blocks for ML rates by about 15% uniformly across every region in January. That covered P5en, P5e, P5 and P4d GPU instances along with Trn2 and Trn1. Jassy said capacity for 2026 is already gone. He said Amazon could invest beyond the roughly $220 billion in capital expenditure planned for this year without it being a poor use of capital, per InfoQ.

The bill arrives in components

Memory is where the capacity crunch turns into a line item. Tom's Hardware reported on 3 November that Apple expects to pay more for memory in the December quarter, citing Reuters. CEO John Ternus called the increases "unprecedented." The company has so far absorbed the shock with inventory bought at lower prices, cost cuts elsewhere and price rises on some products, including Mac and iPad at the end of June and iPhone in September. Tom's Hardware put the rise in its manufacturing bill at $33 billion in a few months.

Apple expects memory purchases to get more expensive into 2027. "For the December quarter, we expect even higher memory costs, partly offset by a few factors," Ternus said, according to Tom's Hardware's account of the Reuters report. Component inflation feeding into device margins is the same supply tightness that is filling data halls.

Governments are treating that tightness as infrastructure. Tom's Hardware reported on 2 November, citing Reuters, that President Vladimir Putin has told owners of non-state data centres in Russia to hold back 10% of capacity for state IT infrastructure on 24 hours' notice. The directive follows Ukrainian drone attacks. In early October drones struck Russia's largest data centre, a 63 MW facility damaged badly enough that modules were taken out of service. Two weeks later the finance ministry's data centre was hit, reportedly home to tax databases and other state registry data. Users reported problems with Yandex Mail, Disk, Documents, Pay and Market after the first strike, though Yandex said key services stayed up.

Where the silicon goes next

The longer argument is about how chips themselves get designed and built. At a science forum in Shanghai on 30 October, Chen Yunji, co-founder of AI chipmaker Cambricon Technologies, predicted that AI and chip design would converge until the design process is automated. "In the [next] 10 years, the most important development ... is the RSI (recursive self-improvement) techniques for artificial intelligence, where AI models improve AI models, and AI chips improve AI chips," the South China Morning Post quoted him saying.

Read together, the week's disclosures describe capacity being added in several places at once. There are module lines in Texas, cell plants in Indonesia, data halls across North America, and pressure inside the memory supply chain that no fab announcement has yet relieved. The US numbers from Waaree are the firmest of the batch. They come with a production start date, a capacity figure and a quarterly revenue line. The rest is either sold out on paper, under construction, or projected to 2030.