Comparing Q2 FY2026 (April 1, 2026) → Q3 FY2026 (July 7, 2026)
“AI-driven demand continued to outpace net sales growth, contributing to a growing backlog.”
That sentence is doing a lot of work.
Penguin Solutions delivered record quarterly net sales of $479 million in Q3, up 48% year-over-year and 40% sequentially. Non-GAAP EPS came in at $0.84 — up 79% from the prior year quarter, up 62% from last quarter. Management raised its full-year outlook for the second consecutive call, this time from 12% net sales growth to 22%, and from $2.15 EPS to $2.60. Then went further, offering a preliminary fiscal 2027 view of approximately 30% growth in both net sales and EPS.
None of that is in dispute. What the transcripts reveal, when read against each other, is something more interesting: a company that delivered substantially above its own expectations but is now managing a set of structural tensions it had, in April, only begun to name.
Delivered
Memory: From Outlook to Blowout
In April’s call, management raised its full-year Memory outlook to 65%–75% growth year-over-year. Nate Olmstead described the revised guidance as driven “majority” by pricing, with strong demand across telecom, networking, and AI-related customers. He flagged one constraint: material availability. “To get to the high end of that outlook really just refers to our ability to secure materials,” he said. “We’re chasing materials.”
Q3 net sales in Integrated Memory reached $275 million — up 111% year-over-year, 60% sequentially. The full-year outlook has been raised again, now to 90%–95% growth. The constraint language around materials has not disappeared; it has been absorbed into a stronger backlog narrative. Demand, in Q3’s framing, is simply growing faster than the company can supply.
The April commitment was directionally correct. The magnitude of the delivery is substantially beyond what the language of that call suggested.
Non-Hyperscale AI Infrastructure: The 50% Promise Becomes 81%
April’s call was specific: non-hyperscale AI/HPC net sales grew 50% year-over-year in the first half of fiscal 2026, representing over 40% of Advanced Computing segment revenue, compared to approximately 20% in the same period last year. Management cited five new logos in Q2, seven for the full first half, against three in the prior year’s first half. The language was growth-affirming but cautious — “still have more work to do,” “typically progress over many months,” “12-18 month sales cycle.”
Q3 delivered non-hyperscale AI infrastructure net sales up 81% year-over-year. The business represented 58% of total Advanced Computing net sales in Q3, versus 33% in Q3 of last year. Four new logos were added in Q3, and management noted that of the 13 new AI infrastructure logos added over the trailing four quarters, seven have already grown their business with Penguin.
The diversification thesis has moved from trajectory to demonstrated pattern.
CXL: From ‘Substantial Order’ to Revenue Line
April’s call announced a “recent substantial order for CXL cards from a generative AI company building solutions for inference workloads” — the first specific revenue signal for the CXL product line. CEO Kash Shaikh framed this as proof-of-concept: CXL adoption was “timely given the transition to inference.”
By Q3, CXL memory expansion cards had generated both revenue and new bookings, the tier-1 financial institution expanded its CXL-based MemoryAI KV cache server purchase for on-premise AI factory deployment, and Deepgram added ClusterWareAI software alongside capacity expansion. The April customer examples have become Q3 expansion examples.
Management also provided hard performance metrics for the MemoryAI KV cache server this quarter: up to 2x higher inference performance, up to 8x lower time-to-first-token latency, and CXL memory approximately 4x–5x more cost-effective than GPU HBM. April’s call had no such figures.
Reprioritized
Advanced Computing Guidance: The Range That Moved in Both Directions
In April, management lowered its Advanced Computing full-year net sales outlook to -25% to -15% year-over-year — a reduction from prior expectations, attributed primarily to deployment timing and the ongoing Penguin Edge wind-down. Kash Shaikh was unambiguous: “The main issue at this point is timing.”
Q3’s revised guidance is -15% to -20% year-over-year. Structurally, that is an improvement at one end and a narrowing at the other — the range has shifted upward but compressed. What changed: non-hyperscale AI infrastructure revenue accelerated meaningfully, partially offsetting the structural headwinds. What has not changed: the Penguin Edge drag, the hyperscale exit assumption, and the combined ~30-percentage-point unfavorable impact language — which appears verbatim in both calls.
Management’s April framing emphasized that bookings were strong but revenue recognition would lag 3–6 months. Q3’s call repeats this framing, now applied to fiscal 2027 visibility. The booking-to-revenue lag is no longer a one-quarter explanation; it has become an ongoing structural feature of how management communicates the business.
Gross Margin: Managed Compression, Harder to Read Than Expected
April’s call explicitly guided for lower gross margins in the second half: “higher mix of lower margin AI hardware and memory sales, rising memory costs in our AI factory solutions, and less tariff cost recovery in LED.” The full-year non-GAAP gross margin outlook was set at 28% ±0.5 percentage points, reduced by one percentage point from prior guidance.
Q3 gross margin came in at 28.1% — described as “above our expectations” despite being 3.6 percentage points lower year-over-year and 3.1 points lower sequentially. Management has now adjusted the full-year gross margin outlook upward to 28.5% ±0.5 percentage points, citing favorable memory pricing in Q3, while explicitly flagging that Q4 will have less pricing favorability.
The framing has shifted from “expect pressure” in April to “beat expectations, expect renewed pressure” in Q3. The trajectory of the underlying margin structure — compressing year-over-year as mix shifts toward higher-volume, lower-margin memory and AI hardware — has not changed. The language around it has become more nuanced.
Share Repurchase Deceleration
April’s call reported $32 million spent repurchasing approximately 1.7 million shares in Q2. Q3 reported $9 million spent repurchasing approximately 466,000 shares. No management commentary addressed this deceleration directly. More significantly, the full-year non-GAAP diluted share count outlook increased from ~53 million shares to ~56 million shares — with Q4 specifically expected to reach approximately 62 million shares due to dilution from convertible debt at higher share prices.
The company moved from active share reduction to net share count expansion within a single quarter.
De-Emphasized or Absent
Georgia Tech: Gone
April’s call named Georgia Tech three times. Kash Shaikh described it as “a strong example” of Penguin’s ability to move organizations “efficiently from concept to production-grade AI infrastructure.” It was paired with Deepgram and Hain as the primary proof-point trio for the diversification thesis.
Q3’s call does not mention Georgia Tech. Deepgram and the tier-1 financial institution both appear — and expand. Georgia Tech does not.
Whether this reflects a change in the relationship, a natural rotation of illustrative examples, or something else is not determinable from the transcripts. What is determinable: a customer named as a strategic validation example in April was not named in July.
Photonic Memory Appliance: Centerpiece to Footnote
In April, the Photonic Memory Appliance occupied meaningful airtime. Kash Shaikh addressed the CXL-to-PMA technology progression in detail during the Q&A, discussing whether photonics needed to work before CXL could scale. The Celestial AI acquisition by Marvell was discussed as a strategic inflection, with $32 million in proceeds received. The PMA was framed as the next frontier.
Q3’s call mentions the PMA in passing — a single paragraph within the product portfolio section. The Marvell-Celestial relationship is acknowledged through “continued partnership with Celestial AI team, now part of Marvell.” No analyst questions addressed it. The technology has moved from a discussed roadmap item to a listed component.
This may simply reflect the relative immaturity of PMA revenue versus the acceleration in CXL and module business. Or it may reflect a shift in internal prioritization. The silence is notable in proportion to the April emphasis.
LED Tariff Recovery: Absent
April’s call identified tariff recovery as a specific gross margin driver for LED. Nate Olmstead named it explicitly, alongside product mix and pricing favorability in memory.
Q3’s LED narrative is compressed into a brief positive update — $66 million, up 7% year-over-year, “managed with discipline.” No tariff mention. The improved LED full-year outlook (now -5% versus the prior -15% to -5% range) implies performance better than feared, but the mechanism has been stripped from the narrative.
Narrative Positioning
‘Durable’ Is No Longer a Hypothesis
In April, management described AI as “adding a more durable layer of demand for memory” — a careful construction. The word “layer” is doing the work: it acknowledges memory’s cyclicality while positioning AI as a structural addition on top of it.
Q3’s language has changed. Kash Shaikh said: “Our strong results reflect a fundamental shift. Agentic AI is driving sustained structural demand for memory, reinforcing our view that this AI-driven demand is more durable than a traditional cyclical memory upturn.”
The hedge is gone. “A more durable layer of demand” has become “more durable than a traditional cyclical memory upturn.” This is not the same claim. The April formulation allowed for cyclicality with an AI overlay. The Q3 formulation positions AI demand as structurally distinct from the cycle. Whether the transcript evidence supports this remains, at minimum, an open question — the memory business also benefited from favorable pricing dynamics that management itself flagged as likely to normalize in Q4.
The Platform Recount
April’s call introduced Penguin’s AI factory platform as built around “six core elements.” They were: Penguin ICE ClusterWare, the new Penguin Memory AI line, Advanced Computing systems, OriginAI factory architectures, end-to-end services, and the partner ecosystem.
Q3’s call describes the platform as having “five core elements with our partner ecosystem” — with the ecosystem now adjacent rather than embedded. The five elements are ClusterWareAI, MemoryAI and integrated memory solutions, advanced computing systems under the “Compute AI” brand, OriginAI factory architectures, and end-to-end services. The rebranding is subtle but present: ClusterWare became ClusterWareAI, Memory AI became MemoryAI, and the Compute AI designation appears for the first time.
None of this was explained.
The Transition Statement
Management opened Q3 by announcing CFO Nate Olmstead’s departure, effective July 8th — the day after this call. The framing was deliberate: “This transition does not change our operating priorities, financial discipline, or focus on execution.”
The statement is structurally defensive — it addresses a concern before the concern is raised. Whether that concern is warranted is a separate question. What is observable: a CFO who delivered two consecutive raised outlooks and managed a complex memory/AI/LED reporting structure is departing mid-cycle, to be replaced initially by an interim, while a formal search begins. The Q3 call was effectively Olmstead’s final earnings call.
Analysts noted it with warmth and moved on to questions about FY2027 guidance. Management’s containment of the narrative held. What Q4 does with the subject will be the more meaningful data point.
What This Quarter Produced
The numbers in Q3 are real. The 104% year-over-year growth in AI-driven businesses, the record operating income, the sequential acceleration across memory and AI infrastructure — these are not narrative constructions. They are delivered results.
What the comparison with April’s call surfaces is a company in the middle of a framing transition. The cautious language of a new CEO finding his footing — “encouraging progress,” “still have more work to do,” the careful hedging around memory cyclicality — has been replaced by more assertive positioning. This quarter’s language is confident in a way the prior quarter’s wasn’t.
Some of that confidence is earned. The non-hyperscale diversification thesis has accelerated faster than guided. CXL has moved from “substantial order” to revenue-generating product. The memory blowout is unambiguous.
The open question entering Q4: whether the confidence is also running slightly ahead of the evidence. Gross margins are compressing. The CFO is gone. Share count is expanding. Georgia Tech has vanished. The PMA has receded. And management’s own Q4 gross margin guidance implies that the pricing tailwind that made Q3 exceptional should not be assumed to repeat.
The preliminary FY2027 guidance — 30% growth — has been introduced into the record. It will need a mirror.
The Gap Report is narrative intelligence, not investment advice.
Quotes are verbatim from publicly available earnings call transcripts. This analysis reflects an interpretation of language and tone shifts between calls
Comments
Log in or sign up to join the conversation.