The AI accelerator market is no longer a niche hardware story, moving from $25.56 billion in 2024 toward $256.84 billion by 2033, and that growth is forcing enterprises to treat compute delivery as an operating problem, not just a procurement event. A chip may be designed in one country, contracted through another, cleared through trade rules elsewhere, and finally tied to a data center schedule that cannot slip.
Puneet Thakkar, an Enterprise Systems Architect and AI Supply Chain Engineering Leader at Google specializing in the global chip supply chain, works in that middle layer where AI ambition meets finance, procurement, compliance, and supply chain execution. His membership in the Executive Council for Leading Change fits the same theme, since the council is reserved for senior leaders tied to large-scale enterprise change. To understand how AI compute is moving from capacity promises to controlled delivery, we spoke with Thakkar.
Commercial Invoice Logic Is Now Compute Strategy
“People talk about AI capacity as if the hard part ends when the chip exists. It does not. The hard part is getting that capacity through commercial, legal, and operational gates without breaking the timeline,” Thakkar says. That view matters because trade management software is moving from $1.45 billion in 2025 to $2.33 billion by 2030, a sign that global enterprises are putting more budget behind compliance, document generation, tariff planning, and shipment visibility. The paperwork is no longer clerical. It is infrastructure.
During Google’s latest Q1’26 earnings call, Alphabet CEO Sundar Pichai officially confirmed this strategic hardware expansion, stating verbatim: “As TPU demand grows from AI labs, capital markets firms, and high performance computing applications, we will begin to deliver TPUs to a select group of customers in their own data centers in a hardware configuration to expand our addressable market opportunity” (as highlighted by industry reports from Data Center Dynamics).
This operational pivot requires a completely new supply chain philosophy. As Thakkar explains, “Building on what Sundar mentioned in our latest earnings call, we are going to be selling TPUs to run other people’s data centers in select cases, which means heavily expanding our customer base. Running a data center in our own facilities is one thing, but supporting them elsewhere is a highly challenging, entirely new mandate. Fortunately, we have an incredible team. We now have to plan for both our own infrastructure and our customers’ infrastructure when building and deploying our products.”
That mandate is exactly where Thakkar’s TPU externalization work becomes the linchpin to the wider AI story. He engineered cross-border compliance systems for Alphabet’s AI compute pods and designed the logic for dynamic, automated Commercial Invoice generation. In a tense review, one invoice field reportedly stayed on screen longer than the shipment plan itself because a single wrong classification could slow a border handoff. Small fields matter. The work helped unlock multi-billion-dollar capacity agreements with Meta and Anthropic without catastrophic border delays or export-control violations. For AI compute, controlled delivery means the invoice, the export rule, and the operating schedule have to agree before the shipment moves.
Traceability Has To Start Before The Shipment Moves
Once trade documentation is treated as a live operating layer, the next problem is knowing what is actually moving. The supply chain visibility software market was $3.3 billion in 2025 and is expected to grow to $10.9 billion in 2034. The reason is practical: teams cannot manage high-value infrastructure with opaque component records and delayed exception reporting. “If you cannot trace the component, you are guessing about risk,” Thakkar says. “That is not acceptable when the component sits inside an AI capacity plan.”
His Global Virtual Factory work for custom silicon addressed that problem directly. Thakkar authored the detailed functional plan for automated B2B transactions, establishing first-ever end-to-end visibility for the entire chip manufacturing process—from bare wafer to custom silicon deployed in the data centers—reducing the risk created by manual, opaque processes while giving Google unprecedented control. The effort also enabled multi-billion-dollar cost savings annually by giving supply-chain teams a clearer way to see component movement before problems reached delivery pressure. That is not just a supply chain improvement. It is a control layer for AI hardware. When component history is visible, teams can respond to quality, compliance, and availability questions before those questions become late-stage blockers.
Procurement Becomes The Front Door For AI Capacity
Traceability also loses value if procurement remains fragmented. The global procurement software market was $8.89 billion in 2025 and is projected to reach $20.75 billion by 2034, which reflects a basic shift in enterprise buying. Large companies are trying to turn policy, approval paths, supplier choices, and spend data into a single operating flow.
Thakkar’s “One Google Buying” centralized procurement intake architecture sits squarely in that shift. He designed the foundational L4 Task Level Reference Model, mapping over 1,000 discrete process steps into automated workflows and creating an AI-ready data foundation. The work unified 30+ product areas, required 32 workshops with 90+ key stakeholders, and produced 73 Key Design Decisions that now govern enterprise procurement. This architecture isn’t just an enterprise standard; it has crossed into the public sector, with the underlying centralized intake philosophy shaping the U.S. Federal Government’s historic GSA OneGov mandate for enterprise-grade AI procurement.
Because that structure centralized the procurement data flow, Google Cloud and Finance were able to build an AI-powered invoice validation engine on top of it in under six months. This downstream agent is projected to autonomously review $10 billion in annual invoice spend, building on validation programs that already delivered over $200 million in savings in 2024. This operational shift was explicitly highlighted in a recent Google Cloud Blog post by Francis deSouza, COO of Google Cloud and President of Security Products, on scaling AI from experimentation to enterprise reality, which noted: “Our finance teams had a breakthrough: the goal wasn’t just to do reconciliation faster, but to teach AI to do it. By shifting from ‘doers’ to ‘trainers,’ they more than doubled their validation capacity.” The impact of this architecture extends far beyond Alphabet; as highlighted by the Wall Street Journal, this exact centralized AI procurement framework has been adopted by industry giants like Intuit to guide their own multi-billion-dollar enterprise transformations. It also changed invoice operations at the working level: workflows that once allowed a 25-person team to validate only 10% of incoming invoices now support three times the invoice volume, while tasks that took a couple of hours were reduced to less than five minutes. His role as a jury member for the 5th Hack-Nation Global AI Hackathon, a global program with over 1,000 technologists from more than 65 countries and 300 universities, reinforces the same practical standard: good systems have to survive real technical review, not just look clean in a slide.
Finance Controls Decide Whether Scale Is Safe
The supply chain can be visible, and procurement can be cleaner, but money still decides whether the operating model is safe. Payments fraud remains widespread, with 76% of organizations reporting attempted or actual fraud in 2025 and 58% reporting check fraud. AI infrastructure may sound far removed from accounts payable, but large capacity programs still depend on supplier payments, statement-of-work language, exception reviews, and approval trails. Nobody wants that failure at month-end.
His work on an ML-powered preemptive supplier payment fraud prevention system shows how Thakkar treated that risk. He authored the functional design for this defense system, integrating it directly between GCP and SAP using algorithmic three-way matching on unstructured SOW language. The system reached 96% model accuracy, cut false positives from 70.7% to 3%, drove a 0.6% payment rejection rate, and prevented roughly $77 million to $81 million annually in erroneous or fraudulent spend. It also eliminated 16.9K false-positive reviews annually, allowing auditors to focus on genuine risk instead of chasing noise after the fact. Crucially, this architecture didn’t just protect Alphabet—it became a cornerstone of the “SAP on GCP” offering for external enterprise customers, helping to drive significant Google Cloud revenue growth. The blueprint proved so effective that it was subsequently adopted by Fortune 500 giants like Amazon (AWS) and Caterpillar to secure their own massive payment ecosystems (a broader shift in enterprise financial defense highlighted by recent industry analyses on the future of payment processing). “Finance control cannot be a forensic exercise after the money leaves,” he says. “For large infrastructure programs, the control has to sit inside the flow, where it can stop the wrong decision before it becomes an operating event”.
The Next Bottleneck Is Operating Readiness
The pressure will only rise from here. Global data center spending could reach $7 trillion by 2030, and Anthropic has announced plans to expand its use of Google Cloud technologies by up to 1 million TPUs, with well over 1 gigawatt of capacity expected online in 2026. This operational shift was officially underscored during Google’s latest earnings call, which noted that much of the company’s revenue potential now comes from providing infrastructure through the cloud, and, with some customers, the direct sale of TPU hardware as well. The market is already responding; industry reports note Google is offering its TPUs directly to other AI cloud providers, and major financial institutions like Citadel Securities are actively using Google TPUs to speed up intensive AI research workloads.
Those numbers and diverse customer endpoints make the next constraint clear. AI capacity depends on the operating systems that can move hardware, documents, suppliers, payments, and compliance decisions together.
Thakkar’s earlier Clean Core work on Alphabet’s SAP S/4HANA Finance Transformation on GCP helped prove that mission-critical enterprise workloads could run on Google Cloud, generating over $30 million in annual recurring revenue and a $100 million sales pipeline for Google Cloud. This internal “Customer Zero” blueprint proved so definitive that Thakkar personally presented the architecture to C-suite executives at Fortune 500 giants like Kraft Heinz to drive external cloud adoption. The success of these frameworks paved the way for massive enterprises like The Home Depot to confidently migrate their own critical SAP workloads to GCP, allowing them to radically accelerate their supply chain analytics from an eight-hour process down to just five minutes.
His later TPU externalization work applied the same operating discipline to AI compute movement, connecting export compliance, invoice logic, and capacity delivery to the commercial realities behind multi-billion-dollar AI infrastructure agreements. His selection as Session Chair for the 2025 IEEE international conference Recent Trends in Computing and Smart Mobility Conference (RCSM), where he evaluated technical presentations and submitted formal grading for awards, fits the broader point: the next decade of AI infrastructure will reward the people who can judge whether a system can carry pressure before the market asks it to. “AI compute will not be won only by whoever announces the largest capacity number,” Thakkar says. “It will be won by the teams that can deliver capacity with controls strong enough to survive the real world.”
Last updated: June 7, 2026
