Penguin Solutions, Inc. (PENG) Earnings

Management highlights

• Kash Shaikh, CEO, starts by thanking Mark Adams for his leadership and mentions AI moving from experimentation to production with workloads shifting to real-time inference. • Builds Penguin into an AI factory platform company with six core elements: Penguin Clusterware, Penguin Memory AI line, Penguin Advanced Computing Systems, Penguin Origin AI factory architectures, end-to-end services, and partner ecosystem. • Addresses second quarter performance: net sales of $343 million, non-GAAP gross margin 31.2%, non-GAAP diluted earnings per share $0.52. • Talks about updated outlook, raising midpoint of full-year net sales and EPS outlook, with integrated memory business fueled by AI-driven demand and advanced computing net sales outlook lower than prior expectations but seeing strong year-over-year Q2 bookings growth for non-hyperscalar AI HPC business. • Details each segment: advanced computing net sales $116 million, integrated memory $172 million, LED $56 million. • Nate Olmstead, CFO, focuses on non-GAAP results, talks about net sales, gross margin, operating margin, earnings per share, balance sheet, and cash flow. • Mentions raising full company net sales and non-GAAP diluted EPS outlook for the year, with advanced computing net sales outlook between minus 25 and minus 15% year over year, memory net sales expected to grow between 65% and 75% year over year, and LED net sales expected to decline between minus 15% and minus 5% year over year. • Talks about non-GAAP gross margin outlook for the full year is now 28%, plus or minus 0.5 percentage points, and non-GAAP full-year diluted earnings per share expected to be approximately $2.15, plus or minus 15 cents.

Guidance

• Raises full company net sales and non-GAAP diluted EPS outlook for the year, with midpoint now calling for 12% net sales growth and $2.15 of non-GAAP diluted EPS, up from previous outlook of 6% net sales growth and $2 of non-GAAP diluted EPS. • Advanced computing: now expects full year net sales to change between minus 25 and minus 15% year over year. • Memory: expected to grow between 65% and 75% year over year. • LED: expected to decline between minus 15% and minus 5% year over year. • Non-GAAP gross margin outlook for the full year is now 28%, plus or minus 0.5 percentage points. • Non-GAAP full-year diluted earnings per share expected to be approximately $2.15, plus or minus 15 cents. • Outlook is based on current environment, including global macroeconomic environment, ongoing supply chain constraints, industry-wide higher costs for memory, etc.

Segment performance

Advanced computing: Q2 net sales were $116 million, representing 34% of total company net sales, down year over year. Non-hyperscale AI HPC net sales grew 50% year over year for the first half of the year, representing over 40% of first half segment net sales. Integrated memory: Q2 net sales were $172 million, representing 50% of total company net sales, grew 63% year over year. LED: Q2 net sales were $56 million, representing 16% of total company net sales, down 7% year over year

Analyst Q&A

• Q: Asks about the raised memory segment outlook for 65 to 75 growth for how much of this is from increased favorable pricing versus demand for new product categories and as a follow-up How should we think about the impacts to the operating margin outlook for this segment and the investments that need to be made into new technologies like CXL and photonic memory appliances.

  A: Nate Olmstead says for the increase in the second half, majority is pricing but demand is also very strong across markets, reflects it in the outlook, expects operating margins for memory to remain pretty healthy in the back half of the year but expects some pressure on gross margins in AI as we see a higher mix of new hardware shipments in the second half, as well as factoring in some of the higher memory input costs in that business.

• Q: In advanced computing what change that caused you to lower the midpoint of your prior guidance to the new range you've communicated? And can you describe how booked you are to that midpoint of that new range?

  A: Kash Shaikh says one of the main factor is the lag between our bookings and the revenue, our revenue lags about three to six months from from the time of the bookings, and given where we are in terms of our fiscal year, we have five months remaining, so going forward, most of the bookings that we are expecting may not materialize into the revenue for the second half of this fiscal, but we believe that it will have A positive impact, obviously, going into the first half of the next fiscal, bookings grew very significantly in Q2 for non-hyperscale AI SPC business, closed five new logos with AISPC in Q2, and in first half, that takes a total to seven new logos as compared to three new logos last year.

• Q: Do the new memory launches mark a shift in strategy on that front? Just curious, because in the past, the company has talked kind of more about the the niche parts of the integrated memory business and noted it's early on things like the CXL front, but now it sounds like memory is expected to be a larger driver as part of this AI factory platform. So just wondering if anything has changed there and what gives you confidence there's durable demand here?

  A: Kash Shaikh says it is a part of our strategy, the memory AI appliances that we launched about a month ago is starting with GPC, is a part of us investing more in our AI factory platform strategy, AI is transitioning from model training to inference, and in the workloads where you are focused on inference, memory becomes an increased requirement, we are leading the market in this area, taking advantage of our unique position at the intersection of memory and AI infrastructure, and one of the new logos we acquired, Tier 1 Financial Institution, purchased our CXL-based KVCash server, which is a proof point of as customers are transitioning from training and bringing AI on-premise in their factories, deploying on-premise, focusing on inference and powering agentic AI.

• Q: With a new CEO in the seat and some moving pieces around sales cycles and supply change, did you change the guidance philosophy at all or embed any additional conservatism? Any color on the puts and takes there would be helpful.

  A: Nate Olmstead says no change in the philosophy, Cash and I very quickly aligned on how we think about tracking the business and looking at things, and with our new CRO, he's done a nice job of adding some more rigor to the planning process in our AI business and just improving the visibility there a little bit, but it's a challenging environment from a supply chain standpoint.

• Q: Double clicks on your advanced computing guidance. You mentioned that a lag of three to six months for the revenue which you will book in your second half but was there a change observed for the for the bookings which you did in first quarter or any change in relative relative to what were you expecting to do in 2q and i have a follow-up as well.

  A: Kash Shaikh says bookings were strong in q2 really good growth sequentially and year over year, do think that the deployment cycle has lengthened a little bit with some of the supply constraints in particular on memory things have gotten a little bit longer, but we're really pleased with the five new logos, and I think demand is good, we're seeing good strength in the pipeline, and it's also diversifying nicely across the non-hyperscale segments such as enterprise and neocloud and sovereign.

• Q: Does NVIDIA coming up with their own reference designs for factory-level solutions, how does that play relative to you? Is that a tailwind for you or is that a headwind for you? Can you please help us understand that?

  A: Kash Shaikh says we believe this is an advantage for us, we work very, very closely with NVIDIA, and some of the wins that I mentioned, for example, the tier one financial institution recently, along with our memory AI product in this transaction, NVIDIA worked very closely with us, and we are working with NVIDIA, leveraging their reference design, combining that with our AI factory platform and complementing NVIDIA's NVI as an example to provide full stack to our customers, so their blueprints are more complementary to our AI factory platform and the components that make up for it.

• Q: A couple, if I could. And maybe Nate as well. So earlier remarks were that you're seeing increased momentum across neoclass sovereign and enterprise. And you mentioned one or two of the new wins. And I think, Cash, you had mentioned you've made some specific or at least general influencing remarks, including around the Gentic. Do you have any specific... context you can give us around what your customers are telling you their thrust in inferencing is right now, and maybe the degree to which Agentsic is showing up there. Like, if we just wanted to get a sense of what the customer's activity tone is, like, behaviorally, say, over the last 90 to 180 days.

  A: Kash Shaikh says we are early in the adoption of inference with these customers but it is increasingly um deployed as in customers as they move towards agentic inference provides the opportunity for powering the agentic and in when when you think about uh when you think about I'll give you an example of why the architecture is changing and why memory is becoming increasingly critical in inference as compared to the model training, for example, let's say if you are writing a book and if you have to write a new sentence without having the memory as a supporting component for you, you will have to reread the entire book before writing the next sentence, so in the inference, you know, you're doing an inference on a lot of data you already have, and if you have a component where the book you have written so far is stored, so before writing the new sentence, you don't have to reread the book, that's kind of how it is changing for the enterprises and other segments, and we see customers already deploying it and architecture is changing, which is why not only we have the opportunity and advantage to provide them our AI infrastructure as well as the services, increasingly we are seeing the demand for our memory AI portfolio where they are deploying AI infrastructure and increasingly inference, they need products like that to be able to provide that memory component for the inference so that the responses of LLMs can be much more faster than they would be otherwise.

• Q: The CXL product, it sounds like you, to the earlier question, it sounds like you guys are a little bit more enthusiastic about the CXL sleeve today than you were maybe 90 days ago. You have the new product out of GTC. Is that accurate statement that you're expecting? Maybe it's because of these new products. a little bit more, and certainly some of the NVIDIA announcements to CES as well. But are you expecting a little bit more revenue a little bit sooner than maybe you were CXL-wise 90 days ago? And then a quick second part to that, do you need photonics to work before you really get CXL amplification? Like, do you need CPO or photonics to work before you can really amplify CXL in CXL? in scale out for scale up thanks that's it for me.

  A: Kash Shaikh says cxl adoption is timely given the transition to inference because as i mentioned with inference you need increased memory for faster llm responses and what cxl provide compute express link is you can share the memory between for GPUs and CPUs, so what it allows is new memory pooling, which is an advantage in inference workloads, while CXL was obviously available for the last, I'd say, few quarters, it is driving that inference adoption is driving the adoption for CXL and this transaction that I mentioned where we received an order It's actually an enterprise generative AI company working on inference workloads, so you can imagine CXL cards make sense for them because those workloads need increased memory and the memory pooling capabilities provided by CXL between GPUs and CPUs are an advantage for those kind of customers, and then in terms of photonic memory appliance that we are working on, in our partnership with Celestial AI, which is now obviously Marvell, that provides increased capability because obviously when you have photonic connectivity, then you have increased capacity to share the memory, so it takes it to the next level, however, CXL in itself is an advantage, we can take it to the next level with the photonic appliance, there is another element, which is KV cache that I mentioned, memory AI KV cache server, which is essentially providing much more responsiveness for larger context workloads, again, used in inference, so various requirements, you can think of it as, inference has various requirements related to memory and the type of workloads it has, and some of it is latency, so these components between CXL or the CXL-based KV cache which provides increased responses and larger memory size, largest context sizes, and then taking it to the next level, photonic memory, make up various use cases for inference.

• Q: The gross margin for the memory, you know, your gross margin was up in the quarter and memory revenue was up strong. And I just want to understand what the dynamics were there.

  A: Nate Olmstead says we saw a little favorability in memory margins, some of that is mixed, a little bit stronger demand than Flash, actually, which is a little bit higher margin product for us within the portfolio, and then also some of the pricing increases, we were able to capture a little bit of margin upside on that just based on the timing of our inventory purchases relative to the timing of, you know, shipments and sales to customers.

• Q: As you get to these CXL systems, would you expect that's going to be a higher margin than the module business?

  A: Nate Olmstead says yeah, we do, it's really a solution, it's got software aspects to it, some good differentiation on the hardware as well, so I see that as a nice margin opportunity for us down the road.