The hardest forecasting problems don't come with data. They arrive as questions: How much inventory do you hold for a product that hasn't shipped yet? How many GPUs do you secure for an AI platform that's still being built? How do you commit capital today for demand that won't materialize for months, or even years?
These are the questions that define modern infrastructure planning, and they've never been more consequential. As artificial intelligence reshapes how companies build and scale, the margin for forecasting error has collapsed. A missed projection on cloud capacity doesn't just mean idle servers; it means delayed product launches, lost customers, and hundreds of millions in misallocated capital. In an era where GPU supply is constrained globally and AI workloads are growing nonlinearly, the organizations that forecast well will scale. The rest will stumble. At scale, these forecasting failures ripple beyond companies—shaping national competitiveness, energy consumption, and the pace at which societies can safely deploy AI.
Amogh Garg has spent nearly fifteen years working at this edge: the space between commitment and uncertainty, where organizations must make resource decisions long before demand becomes visible.
From cloud infrastructure to AI capacity
At Amazon Web Services, Garg worked on programs designed to estimate demand for products that hadn't launched yet. The challenge was fundamental: cloud infrastructure requires procurement commitments months in advance, but new instance types have no usage history to study. Traditional forecasting methods assume you can look backward to project forward. When there's nothing to look back at, you need a different approach—one built on analogous products, feature-based modeling, and cannibalization analysis that estimates how new offerings will absorb demand from existing SKUs while generating net-new usage. These decisions shape not just internal efficiency, but how quickly new technologies reach customers worldwide.
Beyond new product forecasting, he also focused on the uncertainty buffer, which is the gap between a conservative high-end forecast and the average expectation. This buffer exists because forecasters must provision for demand spikes, not just average usage. But an inflated buffer means excess inventory sitting idle; too thin a buffer means stockouts when demand surges. Garg led programs to measure, track, and systematically reduce this buffer, tightening the gap between what AWS expected and what it actually needed, freeing up capital without sacrificing availability.
His work helped AWS make smarter bets on launch footprints and early capacity across complete product lines, improving how capital flowed through one of the largest infrastructure operations in the world.
Today, Amogh leads GPU capacity planning within Salesforce's Cloud Economics & Capacity Management (CECM) organization. Salesforce pioneered bringing CRM fully to the cloud, and is now ushering in the next chapter with Agentforce, its agentic AI platform. Since joining the team in August 2024, he has led the global GPU capacity planning and forecasting essential for the launch of Agentforce, which has now expanded to 16 countries.
The work sits in a space where many variables are still taking shape: demand patterns fluctuate, industry-wide supply remains limited, and AI capabilities are developing faster than traditional forecasting methods can capture. Within this environment, Amogh's focus is on building flexible planning mechanisms that support growth, ensure infrastructure readiness for new model deployments, and retain the agility to adjust or roll back plans when assumptions shift.
A career across supply chains
The instincts behind this work were shaped across an unusual breadth of industries and supply chains.
At Meta, Garg navigated uncertainty of a different kind: measurement systems for advertising that were being rebuilt in real-time after privacy policy changes, such as Apple ATT, disrupted legacy signals. Forecasting in that environment meant building confidence in systems that were still taking shape.
At Microsoft, he brought structure to launch planning across Surface and Xbox, helping the organization move from announcement to availability with greater predictability.
At Reckitt Benckiser, he led demand planning for billion-dollar consumer brands like Lysol and Air Wick, including one of the largest new product launches in the category's history. Decisions about production, distribution, and retail allocation had to be made months before the first unit sold.
At Philips Healthcare, the challenge was spare parts for MRI and CT machines entering markets where no installed base existed. Every stocking decision was a bet on adoption that hadn't happened yet, with patient care hanging in the balance. At the same time, the inventory moved excruciatingly slowly, with individual components sometimes costing upwards of a million dollars each. Overstock meant capital frozen for years; understock meant a hospital's scanner going dark while patients waited for diagnoses. The margin for error was almost nonexistent.
And earlier still, as a Climate Corps Fellow with the Environmental Defense Fund, he built financial projection models for energy efficiency investments, estimating returns on sustainability projects with no operating history, only engineering assumptions and conviction.
From disinfectant wipes to medical imaging parts to cloud servers to AI accelerators—the common thread isn't the product. It's the problem: committing resources before demand becomes visible.
What forecasting for new products actually requires
Garg's experience across these varied supply chains has crystallized into a set of principles for forecasting when historical data doesn't exist:
Start somewhere. Have a forecast; even if it's built on rough assumptions and quick regression modeling, it’s important to get into the game. As actuals accumulate, keep refining. The absence of a forecast does more damage than an inaccurate one. A centralized forecast keeps the whole organization aligned. Without it, every launch manager estimates and procures separately, information stays siloed, and no one learns from misses. With a shared forecast, every signal, big or small, feeds into the system. The numbers become cohesive. The organization understands, collectively, what’s driving projections and why gaps emerge.
Spend time finding and selecting the anchor. In theory, you can forecast without an anchor. In practice, you cannot. Every new product forecast needs a reference point: a similar product from the past, a comparable service from a competitor, an analogous launch with similar characteristics. No matter how sharp your ensemble modeling or how fancy your Bayesian statistics, the choice of anchor can make or break the forecast. A useful technique: build forecasts using multiple anchors and compare the outputs. You would be surprised at the elephantine gaps. Then work on closing them, or consciously choose different anchors for different phases. Pre-launch might reference anchor A; six months post-launch might shift to anchor B as the demand pattern clarifies; later still, anchor C.
Acknowledge that forecasting is more art than science. A model that delivers low MAPE and WAPE at one company may completely fail at another. The same event, whether a product launch, a promotional spike, or a macroeconomic shift, triggers entirely different responses across different supply chains and organizations. What works at AWS may bust entirely in Azure or GCP settings. Context matters more than methodology.
Look beyond the numbers. Low MAPE and WAPE don't guarantee a stable forecast. Sometimes it takes time for customers to realize a product delivers more or less value than they expected, and when that recognition hits, demand shifts suddenly. A newly launched compute instance might show strong price-performance on the spec sheet, yet still require costly re-engineering to migrate existing workloads. That friction won't show up in early adoption curves. Stay close to end customers. Understand whether they're genuinely satisfied or just slow to react. The best forecasters see the reversal coming before it appears in the data.
Embrace imperfection and flexibility. Forecasting isn't about achieving 0% MAPE—it's about striking the right balance between availability and utilization. Chase perfection and you'll either over-provision into waste or under-provision into stockouts. Early forecasts should be structured so they can be revised without sending shocks through the downstream supply chain; if changing a number in week three triggers a crisis in procurement, the system is too rigid for the uncertainty it's operating under. Consider just-in-time procurement during the first six to nine months post-launch. Yes, it costs more, but that premium often delivers more value than capital destroyed by massive early misforecasts. And give yourself time: the intuition to know which teams are chronically over-bullish, which customers consistently underplay their demand, takes months to develop. That gut is as valuable as any statistical model.
The discipline that matters now
Most forecasters wait for data. Garg learned to build the signal first, across supply chains that range from consumer packaged goods to healthcare equipment to hyperscale cloud to AI infrastructure. As compute becomes more constrained and AI workloads more unpredictable, that skill set is no longer niche. It's central to how the next generation of technology companies will allocate capital and scale.
Subject Profile Summary:
Amogh Garg is a leader in infrastructure economics and new product forecasting. With over 15 years of experience, he has spearheaded capacity planning and demand estimation programs at some of the world’s largest technology organizations, including Amazon Web Services (AWS), Meta, and Microsoft. Currently at Salesforce, he focuses on global capacity strategies for the specialized hardware powering the next generation of AI and agentic platforms. His frameworks for forecasting under uncertainty are used to manage complex global supply chains and multi-million dollar capital allocations.
This article was written in cooperation with Julian Monroe