Graphcore
Fabless AI silicon designer (Intelligence Processing Unit + Poplar software stack); post-2024 a captive R&D/compute arm of SoftBank rather than a merchant-market chip vendor - near-zero external product revenue, funded by parent capital toward internal (Stargate) deployment.
Chronological. Series D/E valuations are firm disclosed post-money marks; the 2024 acquisition figure is an undisclosed price reported in a ~$500-600M range (Bloomberg ~$600M), plotted at ~0.55; the 2026 injection is primary capital into the captive subsidiary and did not reset an external valuation.
Earnings, margins, COGS & capex
Financially, Graphcore as a standalone business was a failure - a low-single-digit-million-revenue company burning ~$130-200M/yr, marked down by its own venture backers (Sequoia wrote its stake to zero; Molten cut its holding ~45%). SoftBank bought it for an undisclosed sum reported at ~$500-600M, less than the ~$700M investors had put in, then in Apr 2026 committed a further $457M (described as a 'portion' of expected 2026 funding). The thesis is no longer a P&L story - it is a strategic-optionality bet: cheap access to a proven wafer-scale IPU team, folded into SoftBank's Arm/Ampere 'Silicon Trinity' to feed the Stargate build-out. External commercial traction is effectively nil.
Revenue trend
Margins
loss persisted but narrowed FY2022->FY2023; funded by parent post-2024
COGS structure
Not disclosed. As a fabless designer, cost of revenue centers on TSMC wafer purchases (7nm for GC200, wafer-on-wafer 3D stacking via TSMC SoIC for Bow), packaging/assembly, and system integration; with revenue near zero, COGS is immaterial versus the ~$130-200M annual R&D/opex burn that drives the losses.
Capex
Not disclosed and structurally light (fabless - no owned fab). Spend is R&D/engineering headcount plus SoftBank's announced £1bn Bengaluru engineering campus (~500 semiconductor roles over the next decade) that will absorb capital going forward.
Latest earnings
n/a
No company guidance. SoftBank has publicly framed Izanagi/Stargate 2026 as the deployment milestone; no revenue or unit targets disclosed.
- SoftBank injection (Apr 2026)
- $457M
- Acquisition price (Jul 2024)
- undisclosed, reported ~$500-600M
- Peak private valuation (Dec 2020 Series E)
- ~$2.77B
- Last disclosed revenue
- ~$4M FY2023
Growth drivers
- SoftBank capital — ~$500-600M buyout + $457M Apr-2026 injection removes the funding-death risk that plagued the independent company
- Izanagi next-gen chip — co-developed with Ampere (also SoftBank-owned), fusing IPU parallelism with Arm server CPUs, targeted at SoftBank's Stargate hyperscale data centers in 2026
- Captive internal demand — a guaranteed deployment path inside SoftBank/Stargate rather than fighting for merchant-market share against NVIDIA
- Arm + Ampere + Graphcore vertical integration ('Silicon Trinity') SoftBank controls end-to-end
- Global engineering scale-up (£1bn Bengaluru campus — UK headcount planned to double to ~750, Bristol HQ retained)
Bull & bear
Not a standalone-equity bull case (there's no equity to buy) - the bull case is strategic: SoftBank got a proven wafer-scale AI-silicon team for a reported ~$500-600M, a fraction of replacement cost, and is now pointing it at guaranteed internal Stargate demand rather than the brutal merchant market that killed it. If Izanagi tapes out and deploys, Graphcore becomes the accelerator layer of a fully SoftBank-owned Arm+Ampere+IPU stack - real optionality on a multi-hundred-billion-dollar AI-infra build.
- Acquisition + $457M injection show SoftBank conviction, not a wind-down
- Izanagi (IPU + Ampere Arm CPU) targets Stargate 2026 - a captive, funded deployment path
- Vertical integration with Arm and Ampere is a structural edge no other AI-chip startup has
- Bought cheap (reported ~$500-600M vs ~$2.77B peak / ~$700M invested) - low cost basis on real IP and talent
- Removes the funding-death and go-to-market risks that doomed the independent company
Graphcore already failed once as a business, and acquisition doesn't fix the two things that mattered: NVIDIA's CUDA moat and the absence of external customers. It is now a captive science project whose entire future rests on one unproven chip (Izanagi) hitting an aggressive Stargate timeline for one customer (SoftBank). Its own backers marked it toward zero; hyperscaler in-house silicon has taken the oxygen out of the non-NVIDIA lane; and China is closed. The base rate for independent AI-accelerator startups displacing NVIDIA is ~zero.
- ~$2.7-4M revenue on ~$130-200M/yr losses - never demonstrated it can sell chips
- CUDA ecosystem lock-in makes IPU/Poplar adoption a non-starter outside captive use
- Single-customer, single-product concentration - no commercial fallback if Stargate slips
- VC write-downs (Sequoia to zero, Molten ~-45%) reflect the market's verdict
- Lost networking team to Meta; layoffs and office closures signal talent and market erosion
- Google TPU, AWS Trainium, Microsoft Maia - and now newly-public Cerebras (NASDAQ: CBRS) - already occupy the 'NVIDIA alternative' niche at scale
What it is worth
Private strategic / last-transaction basis - no public equity, negligible revenue, so a revenue or DCF multiple is not meaningful; anchored to the SoftBank transaction and capital commitments plus peak-vs-trough optionality.
If Izanagi slips or Stargate demand disappoints, Graphcore has no external market and value collapses toward the IP/team liquidation floor - consistent with the near-zero marks its own VCs already took.
~$0.5-1B strategic value to SoftBank - a funded captive silicon unit with real IP and a single deployment path; no independent market value.
If Izanagi tapes out and becomes a meaningful accelerator inside Stargate/SoftBank AI infra, the embedded IP + captive-demand position could be worth multiples of the reported ~$500-600M cost basis - a cheap call option on a SoftBank-scale AI-compute build-out.
Last hard marks: an undisclosed SoftBank acquisition price reported at ~$500-600M (Jul 2024), plus a confirmed $457M Apr-2026 injection - implying SoftBank has now committed on the order of ~$1B+ total. Against a ~$2.77B Dec-2020 peak and ~$700M lifetime investment, the buyout was a distressed price. Value today is optionality on Izanagi/Stargate, not fundamentals.
SWOT
Strengths
- Genuinely novel wafer-scale MIMD architecture (Colossus/Bow IPU) with large on-chip SRAM - a real technical differentiator, not vaporware
- Deep, proven silicon + compiler team (founders Nigel Toon, Simon Knowles; ex-Icera/Element14 pedigree)
- Now backstopped by SoftBank's balance sheet - no near-term funding risk
- Privileged access to Arm IP and Ampere CPUs for a vertically integrated stack
- Poplar software stack is mature relative to most non-NVIDIA startups
Weaknesses
- Near-zero external revenue - never achieved product-market fit as a merchant chip vendor
- CUDA lock-in: developers won't leave NVIDIA's software ecosystem for IPU/Poplar
- Own VCs wrote the equity down (Sequoia to zero, Molten ~-45%) - a credibility scar
- Lost its AI-networking team to Meta (2023); repeated layoffs, office closures
- Exited China (US export rules) - cut off from a large addressable market
- Fully dependent on a single customer/parent for demand and capital
Opportunities
- Ride SoftBank's Stargate capex wave as an internal accelerator supplier, sidestepping merchant-market competition
- Izanagi CPU+IPU fusion could differentiate on memory-bound inference vs GPU
- Sovereign/SoftBank-aligned AI-infra demand outside the US-hyperscaler orbit
- Reuse of IPU IP inside broader SoftBank/Arm roadmaps
Threats
- NVIDIA's CUDA moat and relentless cadence (Blackwell/Rubin) keep raising the switching bar
- Hyperscaler in-house silicon (Google TPU, AWS Trainium, Microsoft Maia) crowds the non-NVIDIA lane
- Izanagi execution/tape-out risk on an aggressive 2026 Stargate timeline
- If Stargate demand slips, Graphcore has no fallback commercial market
- US export-control regime continues to shrink its addressable geography
Moats, dependencies & bottlenecks
Moats
real IP, but architectural novelty alone has not translated to adoption Technically differentiated (Colossus GC200, Bow 3D wafer-on-wafer); commercially unproven against GPU economics.
Mature vs peers but decisively behind CUDA; the software gap is the core reason IPUs didn't sell.
SoftBank vertical stack (Arm + Ampere + Graphcore) + captive Stargate demand Moderate-to-strong (as a subsidiary) depends entirely on SoftBank's continued commitment This, not the technology, is the durable advantage post-2024 - guaranteed capital and a captive customer.
Eroded by the Meta networking-team departure and layoffs, being rebuilt via the £1bn Bengaluru campus.
Dependencies
Capital + demand + strategic direction Sole source of funding and effectively sole customer; Graphcore's existence is a SoftBank policy decision.
Foundry / manufacturing All IPU silicon fabricated at TSMC (7nm GC200, SoIC 3D stacking for Bow); no alternative leading-edge foundry.
IP + co-development partner Izanagi fuses IPU with Ampere Arm CPUs; Arm is SoftBank-controlled and Ampere was acquired by SoftBank in 2025, so aligned but concentration-heavy.
Already forced Graphcore's China exit; further tightening shrinks any external addressable market.
Izanagi's payoff hinges on Stargate deployment in 2026; slippage removes the only funded demand.
Advantages
- Cheap cost basis for SoftBank (reported ~$500-600M) on genuinely novel wafer-scale IP and a proven team
- Captive, funded deployment path (Stargate) that bypasses the merchant market that killed it
- Only AI-accelerator player with in-house access to Arm CPU IP + Ampere server CPUs
- Relatively mature software stack (Poplar) versus most non-NVIDIA startups
- Removed existential funding risk under SoftBank ownership
Weaknesses
- Failed commercial track record - ~$2.7-4M revenue, ~$130-200M annual losses
- Software ecosystem far behind CUDA; no developer pull
- Written down by its own venture investors (Sequoia to zero, Molten ~-45%)
- Single-customer / single-product concentration risk
- Cut off from China and much of the export-controlled market
- Talent erosion (Meta departure, layoffs, office closures)
Bottlenecks
- CUDA/software ecosystem lock-in - the single biggest barrier to any IPU adoption
- No external customer base or merchant-market channel to fall back on
- Izanagi tape-out and yield execution on an aggressive 2026 timeline
- Rebuilding engineering depth after Meta took the networking team and post-acquisition attrition
- Total dependence on SoftBank's continued willingness to fund losses
Top signals & trends
Top signals
Concrete commitment; frames Graphcore as an active bet, not a wind-down.
Gives the IPU a funded, captive deployment target - the credible path forward.
Legacy investors' verdict on the standalone business; pre-acquisition scar.
Market-access loss and cost retrenchment before the sale.
Talent stripped from a strategic area.
Scale-up investment signals long-horizon intent.
Trends
Keeps raising the switching cost that blocks IPU adoption in the open market.
Occupies the 'NVIDIA alternative' niche at scale, crowding out merchant challengers.
Creates captive demand for SoftBank-owned silicon - Graphcore's new reason to exist.
Memory-bandwidth-heavy inference could favor IPU+CPU designs like Izanagi if execution lands.
Already closed China; continues to shrink addressable geography.
Ecosystem & competitor graph
Suppliers feed the company; customers pull from it. Line thickness shows the strength of each tie (supply-chain dependency, customer earnings contribution). Hover to isolate a tie.
Sole leading-edge foundry - 7nm GC200, SoIC 3D wafer-on-wafer for Bow IPU.
CPU IP for the Izanagi CPU+IPU design; SoftBank-controlled.
SoftBank-owned Arm server-CPU maker (acquired by SoftBank 2025) co-developing Izanagi. Private - no ticker.
HBM/SRAM and advanced-packaging supply chain (not specifically disclosed).
Parent and de-facto primary customer - internal Stargate/AI-compute deployment.
Early cloud partner/investor (Series D) that previewed IPUs on Azure; limited traction, largely historical.
Historical system/OEM partner and Series-D investor for IPU servers.
Research and enterprise trials historically; never scaled to material revenue.
The incumbent whose CUDA moat and GPU cadence made Graphcore's independent strategy unviable.
Instinct MI-series GPUs + ROCm - the credible #2 merchant AI accelerator.
In-house TPU pods - the most successful non-NVIDIA accelerator at scale.
Captive hyperscaler silicon; the model Graphcore now imitates but at far larger scale.
Gaudi accelerators target the same cost-alternative-to-NVIDIA pitch.
Custom AI-ASIC partner to hyperscalers (e.g. Google TPU, Meta) - takes the merchant-custom niche.
Wafer-scale-engine maker - the closest architectural analog; completed its Nasdaq IPO May 2026 (now public), same 'beat NVIDIA on architecture' thesis but at commercial scale (~$510M FY revenue).
Private LPU inference-focused startup with strong momentum in low-latency inference.
Private AI-accelerator challengers (Tenstorrent led by Jim Keller) competing for the non-NVIDIA lane.
China domestic AI-chip players - context only (US-restricted market Graphcore exited); not a US-investable comparison.