Blaize Holdings, Inc. Common Stock 2026 Q2 Earnings Call

Thank you and good afternoon. With me today are Harminder Sehmi, our Chief Financial Officer, and Stephen Patek, our Chief Revenue Officer. I will start with the outlook and where the business stands. Harminder will take you through the numbers and Stephen will cover our commercial engines. I will then have some closing remarks after the Q&A. As you saw from our earnings release this afternoon, we reduced our revenue outlook for 2026. Our full year revenue is now expected to be between $40 million and $43 million. What that number does not show you is what we have already secured. We hold a signed agreement covering 2,000 servers worth approximately $70 million at current memory prices. Part of that converts into revenue this year. The rest is committed business we carry into 2027. Let me tell you what changed and what did not.

First, several engagements have not converted into orders, including some where pilots were completed successfully. Second, other opportunities are still in progress and expected to close later than we forecast. Third, supply chain cost inflation. Memory pricing has risen materially this year, and we expect that to persist. Harminder will take you through each of them, along with the backlog we expect to be holding at year-end, and what we have changed in how we build our expectations. Behind that number, the business is broadening. Our largest customer in China continues to generate meaningful business for us. We have opened Europe with the first purchase order for several thousand units, and activity across the Asia Pacific region has increased. Let me tell you what we are seeing because it explains both the quarter and the book behind it. The market has made up its mind this year.

Building frontier AI costs more than it earns, and the gap is widening. At the same time, efficient open models are making AI cheaper to run. Value is moving from who trains the biggest model to who runs it most efficiently. The economics of inference are now the deciding factor. That is the market our architecture was designed for, and we are making real progress in it. Two market trends are converging, and we are winning in both. First is physical AI. Countries and industrial companies are putting AI into the field, on vehicles, on ships, on machines, and inside their own borders and their own sites, partly for security and control of their data, but mostly because the work demands it. Speed of response, scale, places the cloud cannot reach. Next is a new generation of AI data centers built to run AI, not just to train it.

Training does not go away. It changes shape into constant tuning and specialization. These sites run many models on many kinds of chips. They are built on purpose, not to depend on one vendor. Both are hybrid, and the operators have learned something important. Renting out GPUs is not a sustainable business. Applications and AI services are. That makes the software that schedules and tunes the work the layer that matters, and that is exactly where we sit. That brings me to what we are focused on most right now, the revenue that we produce and the margin we make on it. Let me take each one. First, revenue. We earn revenue in two ways. One is our silicon and SDK, designed into OEM's product, shipping inside autonomous systems, robotics, and ruggedized equipment. Once we win the design, we scale with that OEM into markets they already serve.

A proven design opens other platforms and markets for us. The other is our hybrid AI platform, a vertically integrated stack. It runs the industry applications that customers buy, built by us and by our software partners. Service providers and system integrators deliver it as a managed service. Stephen will take you through where each of them stands. Second, margin. Behind both sits AI Services, our software suite for AI inference. We expect AI Services to become an increasingly important contributor of our margin over time. Today, AI Services includes capabilities such as facial recognition. Based on requirements from active customer programs, we are developing and integrating additional capabilities, including document processing, quality grading, compliance scoring, video analytics, small language model assistance, and industry-specific services.

We are also developing model optimization and orchestration capabilities intended to route workloads to the appropriate compute resource and optimize models for the underlying hardware. The goal is to give customers more output per dollar of infrastructure. They get more from what they run, and we expect to be paid for what gets them there. We intend to price it as software per rack, per megawatt, or per fleet. That brings me to this quarter. Our gross margin was 8%, reflecting a mix weighted to third-party hardware. Our branded hardware and AI services is what we expect will shift that mix. Building it out is the work in front of us this year. Finally, onto the next generation. The deployments we are supporting today increasingly require a mixture of models and inference workloads.

Serving that demand pulls us deeper into the stack, both in what we build ourselves and what we integrate from others. Based on requirements that we are seeing across current customer engagements, we are working on our next-generation AI inference product designed for production environments. We expect it to complement what we ship today and extend the same architecture to higher-performance workloads. We also intend to incorporate confidential computing capabilities to address requirements from sovereign customers. We view this plan as a staged investment and intend to pace development against customer requirements, commercial progress, and what the business can support. The platform comes first, and the next-generation product is intended to extend that platform into larger inference workloads we expect customers to deploy over the coming years. With that, I will hand it over to Harminder to take you through the outlook and the quarter.

Harminder? Thank you, Dinakar, and good afternoon, everyone.

Before I get into our second quarter results, I will address why we are revising our full year 2026 revenue guidance, what is driving that, and how we are managing the balance sheet through this transition. Dinakar addressed the change from $130 million to a range of between $40 million and $43 million. That is a significant reduction, and I want to walk through exactly why. There are three primary factors behind this change. First, I would like to emphasize that while pilot programs have been successfully completed, several commercial opportunities did not materialize as we expected. We had planned on fulfilling a regular cadence of purchase orders from customers already under contract. With respect to Starshine specifically, we made the decision not to engage further until Starshine pays its outstanding balance.

We have fully reserved the receivable this quarter, engaged local partners to pursue collection, and are re-evaluating that commercial relationship. There is meaningful uncertainty as to whether it will progress further. Second, as we progress into the third quarter, customers are deferring follow-on orders based on their broader scaling of overall deployment of AI solutions. Opportunities are proceeding, just more slowly than our prior forecast assumed. Cloud and data center customers have taken longer to qualify new technology. Certain government programs are on longer procurement timelines than expected, and regional uncertainty has pushed a smart city opportunity in the Middle East into an extended field trial. Finally, memory economics have gotten harder. DRAM and LPDDR pricing has increased materially this year as industry capacity has shifted toward high bandwidth memory. Additionally, the requirement for advanced payments from suppliers has increased. We expect these supply conditions to persist.

Taken together, we have raised the bar for what we're willing to include in guidance going forward. Stephen will walk through the pipeline in more detail shortly, but I want to be precise about how we built this specific number, because the methodology matters as much as the figure. As mentioned, our revised guidance is between $40 million and $43 million for the full year 2026. We project backlog at December 31 this year of approximately $50 million at current memory prices. It is weighted heavily toward revenue from our largest account and based on binding non-cancelable purchase orders that we can fulfill with inventory commitments already made or plan to order. Let me explain the difference between bookings and revenue recognition.