Dragonfly Return

2026-08-26 — Samsung and SK hynix sign on to help build Qualcomm’s memory-stacking answer to HBM

Qualcomm EVP Durga Malladi said at the 2026-08-26 Deutsche Bank Technology Conference that Samsung and SK hynix have agreed to help commercialize Qualcomm’s HBC (High Bandwidth Compute) architecture — which stacks LPDDR DRAM directly on top of a compute die, a near-memory alternative to HBM — after Qualcomm pitched HBC’s advantages and HBM’s shortcomings at its June Investor Day; Malladi said he expected pushback from the two Korean memory makers and instead got an active proposal to collaborate. Qualcomm will design the compute die and TSV integration and work with TSMC on packaging; Samsung/SK hynix handle DRAM stacking. First-generation HBC products are targeted for commercialization/shipment in 2027. HBC is one of the named components of the Dragonfly data-center portfolio (alongside the Dragonfly C1000 CPU and AI300 accelerator), so this is a concrete supply-chain commitment behind the “first silicon dates” question this thread tracks — though it is a memory-partner agreement, not silicon itself, and 2027 is later than the Sept/December contribution timing already on this thread’s radar. (TrendForce, The Elec)

Logged 2026-08-18 by the cold rotation, twenty days late. This thread named Q3 FY26 earnings as its “first live test.”

The $3.9B all-stock Modular acquisition closed, and co-founder Chris Lattner — creator of LLVM, Swift and MLIR — was named to lead Qualcomm’s advanced AI software effort. The anti-CUDA play is now formally staffed and closed rather than pending. Modular’s Mojo language hit 1.0 on 08-11/12, two weeks after the close. (Qualcomm)
On the live test — did Dragonfly move guidance? No. Q3 FY26 revenue ~$9.9B, and on the call Qualcomm reaffirmed rather than raised its datacenter targets: $5B revenue in FY2027, $15B by FY2029. It did name Meta as a Dragonfly C1000 customer — an “Agentic CPU” on the Oryon architecture aimed at agent orchestration and multi-step reasoning. A named hyperscale customer without a guidance change is a real datapoint in both directions.
The open Ventana/RISC-V question resolves as “both, not either.” Ventana’s Veyron V2 custom-silicon business now has two hyperscale customers each contracted for >$1B in FY2027. Dragonfly does not supersede it.
First third-party software-ecosystem move around the new silicon: Multiverse Computing partnered with Qualcomm to optimise models for the AI200/AI250 accelerators, claiming 93% faster performance (08-05). Same day as the Modular close, Qualcomm also signed a 10-year BMW chip-supply deal.

Multiverse Computing and Qualcomm announced a partnership to optimize Multiverse’s compressed AI models for Qualcomm’s Dragonfly AI200 and AI250 accelerators (with Cloud AI100 Ultra used in demos) — aimed at letting data-center operators run more inference on existing hardware at lower power. Multiverse’s Chief Sales Officer Victor Gaspar: “By combining Qualcomm’s efficient AI accelerators with our model compression technology, we can deliver breakthrough improvements in performance, cost, and energy efficiency at scale.” This is an ecosystem/software partnership, not a new hardware customer or a production-date update — a smaller item than the Meta/Microsoft launch-customer news, but the first named software partner since the Investor Day unveiling. (GlobeNewswire via The Globe and Mail)

The Modular acquisition officially closed 07-29 — resolves the thread’s tracked “close expected H2-26” line. Confirmed via Qualcomm’s own newsroom + 15 outlets + an SEC 8-K/10-Q filed same day. Chris Lattner (LLVM/Swift/Mojo creator) named to lead the combined “Advanced AI Software” effort. ([Qualcomm newsroom])
Same day, a separate 10-year BMW automotive-compute deal — auto, not datacenter; adjacent to but not core to Dragonfly’s own watch.
Analyst price-target cuts (Baird/Bernstein/Morgan Stanley/Wells Fargo) are continued reaction to 07-29’s Q3 miss, not new facts.

Q3 FY26: revenue $9.9B beat; EPS $2.21 roughly in-line. The Q4 FY26 (September-quarter) EPS guide of $2.05-2.25 missed consensus (~$2.38), driven entirely by the legacy business — memory/input-cost inflation and an accelerated Apple-modem revenue decline, now seen “materially lower” than the prior ~20% estimate. (CNBC)
Dragonfly did not move this guide. The two custom-silicon engagements are set to start contributing revenue in the December quarter (fiscal Q1 FY27) — the quarter after this one. (CNBC)
The AI story moved on a longer horizon instead: FY2029 non-handset target raised to $40B from $22B, including more than $15B in data-center revenue; HBC Gen 1 taped out, targeting a mid-2027 launch. (GuruFocus)
Stock fell regardless of the improved long-range story — -8% after-hours, -4.8% premarket — trading tonight’s legacy-business weakness rather than 2029’s promise.

Dragonfly C1000 CPU + AI300 accelerator launched — Meta and Microsoft named launch customers, $15B datacenter revenue by 2029 target; stock +15%. Same day: acquired Modular ($3.9B) — an explicit anti-CUDA inference-software play. Analysts frame it as the return to the market Qualcomm exited after Centriq (~2018).

The reported $8-10B Tenstorrent talks were real (06-15), then denied by Tenstorrent’s CEO (06-29) — no close. Ventana’s RISC-V line status unclear: superseded by Dragonfly? (open gap)