Compute & Infrastructure

CG Power's semiconductor unit, CG Semi, has begun production with an announced annual capacity of 200 million chips, a development Prime Minister Modi framed as a milestone in India's sovereign semiconductor strategy. Bloomberg reported the announcement on July 4, citing Modi directly. The facility represents the first confirmed production output from India's $10 billion-plus incentive program to build a domestic chip industry, moving the country from policy announcement to operational reality.

Strategic context is essential here. 200 million chips annually is a meaningful starting point for compound semiconductors or legacy-node chips — the type of devices used in power management, automotive, and IoT applications — but it is orders of magnitude below what TSMC produces in a single month across its advanced nodes. India's near-term play is not to challenge Taiwan on leading-edge logic; it is to capture a defensible position in the mid-range supply chain and reduce import dependency in sectors the government deems strategic. The more important signal is whether this facility reaches consistent yield targets and whether it catalyses follow-on investment from global IDMs seeking China-alternative packaging and test capacity in Asia.

TrendForce data cited by Tom's Hardware shows DRAM and NAND prices continuing to climb through Q3 2026, but with a structural split emerging. Enterprise DRAM — particularly HBM3E used in AI accelerators — remains supply-constrained and price-inelastic, driven by hyperscaler and cloud infrastructure procurement. Consumer-facing DRAM and NAND, by contrast, is hitting affordability ceilings as PC and smartphone OEMs resist passing further cost increases to end users.

This bifurcation has significant supply chain implications. Samsung, SK Hynix, and Micron have all been shifting wafer allocation toward HBM and high-capacity server DRAM, which commands substantially higher margins. The risk is that if AI infrastructure spending decelerates — whether from a macro shock, export control tightening, or a pause in hyperscaler capex cycles — the memory majors will be caught with capacity weighted toward a segment that drops sharply, while consumer demand has been underserved and inventory restocking takes time. For infrastructure buyers, the practical consequence is that commodity memory procurement costs will remain elevated through at least Q3 even as the AI-specific premium component stays supply-constrained.

Together AI's $800 million Series C, valuing the company at $8.3 billion post-money, is the largest disclosed funding round for an AI inference cloud provider in 2026. Data Center Dynamics reported the raise without disclosing lead investors, which limits independent verification of the valuation basis. Together AI operates as a managed inference layer — renting GPU clusters from hyperscalers or co-location providers and reselling optimized inference access to developers — placing it in the increasingly crowded middle tier of the AI compute stack.

GMI Cloud CEO Alex Yeh's appearance at IVS2026 in Kyoto, covered by Bloomberg, adds a regional dimension: GMI operates GPU cloud infrastructure with a focus on Asia-Pacific markets where hyperscaler capacity is tighter relative to demand. Together these datapoints confirm that the market for GPU cloud intermediaries — companies that aggregate scarce compute and resell it with value-added tooling — continues to attract large-scale capital. The structural question is whether this tier survives as hyperscalers build out their own managed inference products and compete directly on price.

Startup Ampera has unveiled what it describes as the world's first subcritical, solid-state, factory-built thorium reactor module manufactured using 3D printing techniques, explicitly targeting AI data centre power supply. Tom's Hardware reported the announcement. Critical caveats apply: subcritical reactor designs do not sustain a chain reaction independently and require an external neutron source, which materially affects their net energy output profile. No regulatory submissions, independent engineering reviews, or pilot deployment timelines have been confirmed.

The announcement is nonetheless analytically significant as a demand signal rather than a supply signal. The fact that a startup is raising capital on a data centre power narrative — even one built on speculative hardware — reflects how acute and widely understood the grid capacity constraint has become. Confirmed nuclear-adjacent infrastructure plays in the data centre sector remain limited to Microsoft's restart of Three Mile Island Unit 1 and a handful of SMR offtake agreements with NuScale and X-energy, all of which are years from delivering power. Ampera's thorium module belongs in the speculative-concept category until regulatory engagement and third-party technical validation are confirmed.

Legrand has delivered an NVIDIA-powered supercomputer to Yerevan State University in Armenia, installed in retrofitted university buildings, according to Data Center Dynamics. While modest in global scale, this deployment illustrates the broadening geographic distribution of NVIDIA GPU infrastructure beyond the G7 compute powers. Armenia sits in a strategically sensitive corridor between Russia, Turkey, and Iran, and university-anchored supercomputer deployments in such locations typically serve dual purposes: building domestic AI research capacity and establishing a compute sovereignty baseline.

The use of Legrand — a French electrical infrastructure and data centre equipment company — as the delivery and integration partner rather than a US hyperscaler or direct NVIDIA channel is notable. It suggests that mid-market sovereign compute deployments are being served through European system integrators, potentially as a way to navigate US export control sensitivities. The retrofit installation in existing university buildings rather than purpose-built data centre infrastructure also points to the pragmatic, resource-constrained nature of this tier of sovereign compute investment.