L&T's $1.2 Billion Chennai AI Factory Reflects Nvidia's Widening Global Compute Reach
Larsen & Toubro's contract to build India's largest Nvidia B300 AI factory is the latest indicator of a Nvidia-centric global buildout that is simultaneously testing supply chains, reshaping capital markets, and attracting new geopolitical risks.
When Larsen & Toubro, India's largest engineering conglomerate, announced on August 14 that it had secured what it classified as a 'mega order' — reportedly worth approximately $1.2 billion (₹10,000 crore) from Together AI — to build India's largest Nvidia B300 AI factory in Chennai, the news drew swift attention. But the deal's significance extends well beyond Indian infrastructure ambitions. It is the latest data point in an accelerating, Nvidia-centric buildout that is rewiring global compute geography and testing the limits of both supply chains and financial engineering simultaneously.
The Chennai facility will house more than 10,000 Nvidia B300 GPUs, according to Times of India reporting, a milestone for South Asian AI infrastructure that arrives in the same week Together AI announced a separate $240 million deal with IBM to deploy Nvidia HGX B300 clusters on IBM Cloud targeting regulated enterprise workloads. Placing simultaneous bets across hyperscalers, sovereign clouds, and specialized inference operators, Together AI's strategy reflects the demand confidence that has now pulled Wall Street directly into the compute funding chain — and illustrates how Nvidia's architecture has become the common denominator across otherwise competing deployment models.
Nvidia's own role in that chain expanded sharply this week. The company announced a $500 billion AI infrastructure financing program structured with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR — positioning Nvidia as a capital intermediary alongside its core identity as a chip designer. The arrangement drew immediate scrutiny: CNBC reported that investors are questioning data center loan valuations following the announcement, a signal that the financial engineering underpinning the buildout faces growing examination even as orders accelerate. Separately, IREN secured a $3.4 billion, five-year AI cloud contract with Nvidia and became the first operator to receive Nvidia's newly created 'Exemplar Cloud' designation — tied to its GB300 deployment at Microsoft's Horizon 1 — indicating that Nvidia is building a formal quality tier among its cloud operator partners, with hardware access likely to follow.
The hardware roadmap reveals both ambition and constraint. Nvidia's next-generation Kyber platform — the rack-scale system housing Rubin Ultra GPUs — is reportedly priced at $41.6 million per rack according to Wccftech, with each unit carrying 340.4 terabytes of DRAM and HBM4E memory priced at approximately $19.76 per gigabyte. The target specification for Rubin Ultra is 768 gigabytes of HBM4E per GPU. Yet Tom's Hardware reported this week that Nvidia has been testing configurations as low as 192 gigabytes using standard HBM4, a sign that memory supply constraints are already forcing design flexibility. SK Hynix, Samsung, and Micron are competing for Q4 HBM4 delivery contracts. TSMC — which reported 45% year-over-year revenue growth driven largely by Nvidia, AMD, and Apple AI chip demand — is running at elevated utilization, but memory, not logic, is the binding constraint in the near term.
Two longer-horizon risks deserve equal weight alongside the opportunity narrative. First, China: analysis published this week suggests that domestic Chinese AI chip development, accelerated by U.S. export controls, could capture 80 to 90 percent of China's AI market within a few years, structurally reducing one of the largest potential demand pools for Nvidia hardware. Second, the financial architecture itself: the $500 billion program effectively uses data center assets as collateral for a new class of AI infrastructure loans, and investor skepticism about underlying valuations — already surfacing in public commentary — could constrain the credit available to the very operators placing today's orders. CoreWeave's second-quarter revenue of $2.58 billion, up 112 percent year over year, and its A100 contracts extending to 2029 confirm that legacy hardware remains commercially viable, providing some floor to operator economics — but they also illustrate how capital-intensive and long-dated these commitments have become.
For investors and operators tracking Nvidia, three signals are worth close attention over the coming quarters: whether Kyber rack shipments begin at the full 768-gigabyte HBM4E specification or at reduced configurations, which would indicate the real severity of memory supply tightness; the pace at which Nvidia's Exemplar Cloud certification program expands, as it appears to function as a mechanism for tiering future hardware access among operators; and the terms and early performance of the first wave of AI infrastructure loans to emerge from the Wall Street financing program, which will determine whether the capital formation story holds together as the buildout matures.