Every investment blog writes retrospective analysis — explaining failures after the fact. That’s easy. It’s also useless, because the explanation always fits when you already know the answer.

We do something different. We run our multi-agent Analytical Council on the company using only information available at the time of investment, lock the verdict, and then compare it against reality. This is a blind backtest — the same methodology we use to calibrate our system across 200+ deals. Here’s Graphcore.

Phase 1: The Blind Assessment (December 2020)

We fenced our system to December 2020 — the date of Graphcore’s $222M Series E at a $2.77B valuation. No knowledge of anything after that date. Three analytical lenses assessed the deal independently, then our Associate synthesised.

Technology & AI Assessment flagged ecosystem lock-in as the binding risk. The IPU architecture was genuinely differentiated — real performance gains on sparse and graph-based workloads. But GPT-3’s release six months earlier was a negative signal: dense transformer scaling was becoming the dominant paradigm, moving the market away from Graphcore’s strengths. NVIDIA’s A100 (May 2020) showed the incumbent was closing architectural gaps faster than Graphcore could build ecosystem.

Returns & Unit Economics computed a 277x revenue multiple ($2.77B on ~$10M FY2019 revenue). Burn rate of $110-150M/year left roughly 3 years of runway. Base case return: 2-3x over 5-7 years via acquisition. The channel partner pivot (Dell, Sep 2020) was noted as a possible early signal that direct sales was insufficient.

Operations & Execution gave the highest marks to the founding team. Nigel Toon and Simon Knowles had two prior exits including Icera to NVIDIA for $435M. On-time GC200 delivery was a strong signal in hardware. But scaling a new architecture against NVIDIA’s 1,000-person operations team with entrenched customer relationships was identified as the core execution challenge.

Locked verdict: Don’t invest in this round, but don’t write it off either.

The technology is real, the team is exceptional, but the market is moving against them. Re-evaluate in 12-18 months if: (1) revenue reaches $30-50M annually, (2) customer base expands beyond 10 enterprise accounts, (3) their software framework reaches feature parity with the NVIDIA ecosystem, (4) NVIDIA’s next-gen chip does NOT address the specific workloads where Graphcore has an edge.

Phase 2: What Actually Happened

SoftBank acquired Graphcore in July 2024 for approximately $600M — a 78% discount to the $2.77B peak valuation, on just $4M of annual revenue. The IPU architecture failed to gain ecosystem traction. Revenue didn’t grow from $10M; it declined to $4M by FY2023.

Our system’s grade: directionally correct. The blind assessment said “don’t invest” and the company lost 78% of its value. The caution was warranted. But the reasoning wasn’t perfect — more on that below.

All four re-engagement conditions were missed: revenue collapsed rather than growing, customer count didn’t expand meaningfully, the software framework never caught up, and NVIDIA’s H100 (2022) and Blackwell (2024) further entrenched GPU dominance.

Phase 3: What We Learned

What the blind assessment got right: The GPT-3 signal — identified as negative for Graphcore in the blind assessment — turned out to be the most important call. Dense transformer dominance accelerated from GPT-3 to ChatGPT to GPT-4 in three years, completely locking AI compute into workloads optimised for GPUs. The ecosystem lock-in thesis was correct. The team quality assessment was also correct: Toon and Knowles’s track record prevented a total shutdown (SoftBank acquired for talent), confirming that exceptional teams prevent death but don’t prevent value destruction.

What the blind assessment got wrong: We assumed revenue would grow monotonically — predicting $30-50M by Q4 2021. Revenue actually declined. Neither version modeled the possibility that revenue could go down, not just plateau. This is a systematic gap we’ve now added to our calibration: for companies dependent on ecosystem adoption that hasn’t materialised, revenue decline is a real scenario, not just stagnation.

The GPT-3 signal, while correctly identified, was under-weighted. The assessment treated it as one risk among several, when in hindsight it was the dominant one. The system should have been more bearish than it was. We’ve added a calibration rule: when a fundamental platform shift is underway during the assessment window, weigh it more heavily.

What neither could have predicted: The speed of transformer dominance. In December 2020, dense transformers were one of several promising ML paradigms. By 2023, they were the only paradigm that mattered at scale. SoftBank as the acquirer was also unpredictable — their chip ambitions (post-ARM) created a buyer that didn’t exist in the typical semiconductor acquirer set.

The Charaka View

This is how multi-agent analysis compounds. One assessment produces a verdict, a score against reality, and four calibration rules that make the next assessment better. A single analyst might have called Graphcore correctly — many did. The difference is that our system records why, checks itself, and updates its priors. After 200+ blind assessments, the patterns become durable: ecosystem lock-in kills differentiated technology, exceptional teams prevent death but not down-rounds, and revenue monotonicity is an assumption that must be tested, not assumed.

Our Methodology

Every Death Diagnosis article follows this structure. We run our multi-agent Analytical Council using only information available at the assessment date — no hindsight, no retrofitting. The verdict is locked before we look at the outcome. Then we score it, identify what we got right and wrong, and feed the learnings back into our calibration system.

This is the same blind backtest methodology we use across our full pipeline of 200+ assessed companies. Our current weighted accuracy is 66.7% — not perfect, and we publish that number because transparency about what we get wrong is more valuable than pretending we get everything right. Every Death Diagnosis makes the next one more accurate.

We believe this is the only investment content format that is genuinely falsifiable. If you know of another, we’d like to hear about it.

This analysis was generated by Manthan Intelligence’s analytical system — a continuously growing knowledge graph, multi-agent Analytical Council, and calibrated scoring methodology. Human editorial oversight applied.

Charaka Notes by Manthan Intelligence. Subscribe