10-K discloses $134 billion of RPO, NVIDIA standardises NVHBM specification on TSMC 3nm across the board…
Micron's FY2026 10-K discloses remaining performance obligations (RPO) of about $134 billion for the first time (only $5 billion at the end of the previous quarter). The bulk comes from take-or-pay agreements with strategic customers, which give a real cushion in a downturn through price floors and locked-in volumes.
Read more
But 75% falls after 12 months, and near-term revenue still moves with spot and contract prices. NVIDIA requires that the NVHBM base dies of Samsung, SK Hynix and Micron all use TSMC 3nm, and that the memory core dies all use the sixth-generation DRAM process. Samsung is giving up its one-stop in-house foundry model for the first time, so the design moat is diluted and competition returns to yield, capacity and delivery. Micron's monthly HBM capacity is only about a third of the Korean makers', so it remains capacity-constrained. Lai Ching-te has called for sincere negotiation and affirmed Micron's investment commitments in Taiwan, which lowers the probability of the labour dispute escalating. The second mediation session in Taichung on 22 Oct is the key date.
Key evidence
- Micron filed its FY2026 10-K with the SEC on 9 Oct (XBRL checked): the transaction price allocated to remaining performance obligations was about $134 billion as of 3 Sep. At the end of the previous quarter, 28 May, it was only $5 billion, and at the end of FY2025 it was immaterial.
- RPO structure: the strategic customer agreements are take-or-pay with binding locked-in volumes, and most set a price range between a floor and a cap. Of the total, $13 billion is already booked as contract liabilities (mainly customer deposits, almost all scheduled to be refunded in 2029 to 2031).
- RPO timing: about a quarter (about $33.5 billion) is expected to be recognised as revenue in the next 12 months, and all of it within five years. Against FY2026 revenue of $133.19 billion, the contracts lock in about a quarter of a year's revenue for next year.
- The company states explicitly that RPO is based on locked-in volume multiplied by minimum price and does not represent future actual revenue. 75% falls after 12 months, and near-term revenue still moves mainly with spot and contract prices.
- Seoul Economic Daily, 18 Oct: NVIDIA requires Samsung, SK Hynix and Micron to use TSMC 3nm for the base dies of NVHBM (based on the HBM4E architecture) across the board, with memory core dies standardised on the sixth-generation DRAM process (1c / 1γ class).
- NVHBM design: the memory controller moves down from the GPU to the base die, the physical-layer circuitry shrinks by 60%, bandwidth is up to 30% higher than standard HBM4E and power consumption is 15% lower.
Show 5 more
- Samsung accepts a TSMC 3nm base die plan for the first time, giving up its one-stop in-house foundry model for NVIDIA (it keeps it for other customers such as AMD).
- Micron's monthly HBM capacity is only about a third of the Korean makers', so the ceiling on its share is still capacity, not design.
- UBS forecasts 2027 HBM demand of about 61.5 billion Gb, up 91% YoY, against new supply of about 54.6 billion Gb, a shortfall of about 10%.
- Potential bottleneck: all three makers' base dies are crowded onto TSMC 3nm (TSMC is also expanding 3nm capacity at Fab 18 in Tainan, Arizona and Kumamoto). Samsung is also qualifying its own 2nm as a hedge.
- Taipei Times, 18 Oct: Lai Ching-te called on Micron and the union to negotiate sincerely, and affirmed Micron's commitment to Taiwan and its continued expansion of investment there. The second mediation session in Taichung on 22 Oct is the key date.
Catalysts to watch
- Second labour mediation session at Micron Taichung on 22 Oct
- FY2026 Q1 results (changes in RPO, the trend in contract liabilities, how the HBM capacity ramp is reported)
- How NVIDIA's NVHBM volume orders are actually split among the three makers (whether Micron can use its existing partnership with TSMC to win share)
- The pace of TSMC's 3nm capacity expansion (Fab 18 in Tainan, Arizona, Kumamoto), which determines the supply bottleneck for NVHBM base dies
- Whether the roughly 10% HBM supply-demand gap in 2027 materialises
Risks
- RPO is based on locked-in volume multiplied by minimum price and does not represent actual revenue. Near-term revenue is still exposed to spot and contract price swings
- Once NVHBM specifications are standardised, NVIDIA can switch orders between the three makers more flexibly, which dilutes the moat from custom HBM lock-in
- Micron's HBM capacity is only about a third of the Korean makers', so it cannot win share on design strength alone
- All three makers' base dies are loaded onto TSMC 3nm, so if 3nm capacity is tight, deliveries by all three could suffer at once
- If the labour mediation in Taichung on 22 Oct breaks down, the dispute could still escalate
Sources (4)
- Micron FY2026 10-K (SEC, 9 Oct, checked against XBRL)
- Seoul Economic Daily, 18 Oct
- UBS forecast of 2027 HBM supply and demand
- Taipei Times, 18 Oct
A brief is a lead for our research, not a verified conclusion or advice.