Case Study: Semiconductor Market Sizing: Reconciling a 3× TAM Gap

This market sizing analysis case study shows how PerspireIP builds a defensible total addressable market for a semiconductor company when the published figures disagree with each other by threefold. Sizing a fast-moving chip sub-segment is where market models most often fail: third-party reports are stale, definitions differ, and a founder can find a number to support almost any story. The worked example below is representative — assembled from public market and patent data — but the bottom-up method is exactly the one we run on a live engagement.

The Challenge

A venture-backed company held packaging and thermal-interface IP that attaches to high-bandwidth memory (HBM) stacks inside AI accelerators — roughly 40 granted and pending patent families. It was raising a Series B, and the deck led with a total addressable market lifted from a single market report. The trouble was that a second report, equally reputable, put the same 2030 market at less than half the figure, and a third put it far higher.

An investor’s analyst had already flagged the discrepancy. When a headline TAM can be tripled or halved by swapping which report you cite, the number stops persuading anyone. The company asked us for a market sizing analysis case study of its own segment — a bottom-up model it could defend line by line in the data room, with the divergence between the published figures explained rather than hidden.

Our Approach to This Market Sizing Analysis Case Study

This market sizing analysis case study ran through three proven steps. Each answers a question the previous one leaves open, and together they turn three contradictory report headlines into one number a founder can defend:

1. Step 1 — Reconcile the published figures on one definition. Restate every third-party forecast against a single, explicit boundary for what counts as the market, so the reports become comparable before any of them is trusted.
2. Step 2 — Build the market bottom-up from an observable unit. Derive the total from AI-accelerator shipments × HBM stacks per accelerator × content value per stack, so every input is a number an investor can independently check.
3. Step 3 — Corroborate and layer TAM, SAM and SOM. Cross-check the bottom-up total against the reconciled band and the patent-filing signal, then carry the assumptions forward into a serviceable and obtainable market.

We refused to start from any single published headline. The observable anchors were public: the global semiconductor market reached $627.6 billion in 2024, up 19.1% year over year, and WSTS forecasts it surging past $1.5 trillion in 2026, with the memory segment projected to more than triple as HBM feeds AI demand. Within that, the HBM market itself was roughly $2.9 billion in 2024.

The bottom-up driver was the accelerator. Public teardowns show each leading AI GPU — NVIDIA’s H100 and B200 class — pairs the compute die with six to eight HBM stacks, and large-language-model training runs consume tens of thousands of GPUs. So we sized the market as accelerator units shipped, multiplied by stacks per unit, multiplied by content value per stack — a chain in which every link is a number an investor can check, rather than a market-share percentage asserted against a headline.

Keeping the three layers separate mattered as much as the arithmetic. The TAM is the whole HBM market the client’s packaging IP could theoretically attach to; the SAM is the slice reachable given its process, customer set and patent rights; the SOM is the share realistically winnable over the funding horizon. A TAM quoted as if it were revenue is exactly the mistake the original deck made, and the one an analyst spots first.

What the Research Found

Three findings shaped the reconciled range:

• The divergence was a definition problem, not a data problem. The published 2030 HBM forecasts we compared ranged from about $9.8 billion to over $30 billion — roughly a threefold spread — mostly because each drew the boundary of “HBM” differently (bare stacks vs. stacks-plus-packaging vs. HBM-enabled modules). Restating them on one definition collapsed most of the gap.
• The bottom-up total landed inside the reconciled band. Building from accelerator shipments × stacks per unit × content value produced a TAM that sat between the conservative and aggressive third-party figures — corroboration from an independent method, which is what an analyst actually trusts.
• The filing signal supported the demand story. WIPO recorded 3.7 million patent applications worldwide in 2024, up 4.9% and a fifth straight year of growth, with computer technology the single largest and fastest-growing field of the decade — evidence the client’s area was compounding, not cresting.

Where our bottom-up model and the reconciled third-party band overlapped, the client had a TAM it could defend; where they diverged, we documented exactly why, which is what a sharp investor looks for first.

The Outcome

The market sizing analysis case study replaced a single contestable headline with a bottom-up model whose every input traced to a cited source, plus a one-page reconciliation showing why the published reports disagreed. In the scenario, that turned the analyst’s objection into a strength — the founder could now walk the data room through the number instead of defending a citation.

The SAM layer did double duty: it became the go-to-market narrative (which accelerator platforms the client’s packaging IP could realistically attach to), while the SOM layer set a credible revenue ramp for the model. One market model, consistent from the pitch narrative through to the financial projections, with no drift between the story and the spreadsheet.

What This Means for Similar Matters

The pattern generalises to any market sizing analysis case study in a fast-moving field:

• Never lead with one report’s headline in a category where reports disagree — reconcile them on a single definition first.
• Build bottom-up from an observable unit (here, accelerators × stacks × content value); a derived total beats an asserted market share every time.
• Use published figures as a sanity ceiling, and treat any order-of-magnitude gap as a finding to explain, not a number to pick.
• Log every assumption. A TAM an investor can inspect and still not break is worth more than a bigger one they can dismantle in a single question.

A market sized this way does not just survive diligence — it shifts the burden to the skeptic to explain why a fully-sourced, independently-corroborated number is wrong.

Why Fast-Moving Chip Markets Resist Sizing

Semiconductor sub-segments are among the hardest markets to size, and the reasons are structural rather than a failing of any one analyst. Product generations turn over in eighteen months, so a report researched a year ago already describes a different market. Definitions are contested — one house counts bare memory stacks, another counts the packaged module, a third counts the system it enables — and each choice moves the total by billions. And demand is concentrated in a handful of buyers whose roadmaps are secret, so top-down share assumptions are little more than guesses dressed as data.

That is precisely why a bottom-up model wins in a data room. When the unit of demand is an accelerator that ships in observable volumes and carries a countable number of memory stacks, the analyst can rebuild the total from first principles and check it against the reconciled reports. The method does not pretend the uncertainty away; it makes the uncertainty explicit and bounded, which is the most an honest market model can promise in a field moving this fast.

How the Market Number Fed the Financing

A sized market is only useful if it drives a decision, so we built the model to serve the raise directly. The reconciled TAM set the ceiling of the opportunity narrative; the SAM — the accelerator platforms the client’s packaging IP could realistically attach to — became the addressable-market slide investors actually underwrite; and the SOM, tied to a credible ramp of design wins over the funding horizon, anchored the revenue line in the financial model.

Because all three layers came from one traceable model, the pitch narrative and the spreadsheet could not drift apart — a common failure where the market slide claims one story and the projections quietly assume another. In the scenario, that internal consistency was what let the founder move the conversation from “which report do we believe?” to “here is how we capture our share,” which is the only market question a term sheet actually rewards.

Data Sources

The market and patent data referenced above comes from:

• Semiconductor Industry Association — Global Semiconductor Sales 2024 — Global chip sales $627.6B in 2024, up 19.1% on 2023's $526.8B
• WSTS — Semiconductor Market Forecast — Market forecast past $1.5T in 2026; memory segment surging on HBM/AI demand
• WIPO — World Intellectual Property Indicators 2025 (Patents Highlights) — 3.7M patent applications in 2024 (+4.9%); computer technology the fastest-growing field
• Precedence Research — High-Bandwidth Memory Market — HBM market ~$2.9B (2024); one of several diverging long-range forecasts reconciled

Discuss a Market Sizing Engagement

Turn a contestable headline into a number your investors and licensees will accept.

Related PerspireIP work: Market Sizing & Opportunity Analysis · Technology Scouting · IP Valuation.

Frequently Asked Questions