Sandisk HBF Roadmap: Targeting 2027 Samples to Work Alongside HBM4

AI Memory Roadmap · HBF NAND · HBM4 · Datacenter Infrastructure

Sandisk HBF Roadmap: Targeting 2027 Samples to Work Alongside HBM4

A comprehensive analysis of how Sandisk’s high-bandwidth flash roadmap could align with HBM4-era AI accelerators, reshape memory tiering, and influence the economics of future AI datacenters.

1. Executive Summary

The AI infrastructure market is entering a period in which memory architecture may determine the economics of large-scale deployment as much as raw accelerator performance. With larger foundation models, longer context windows, retrieval-augmented systems, multimodal applications, and AI agents becoming mainstream, datacenters are under pressure to deliver more memory bandwidth, more memory capacity, lower latency, and better energy efficiency at the same time. High Bandwidth Memory, or HBM, has become the central premium memory technology for AI accelerators. HBM4, expected to define the next major performance tier, will push bandwidth and integration further for advanced training and inference platforms.

Against this backdrop, Sandisk’s HBF roadmap, reportedly targeting samples around 2027 to work alongside HBM4, deserves close attention. HBF, broadly understood as High Bandwidth Flash, represents an attempt to reposition NAND flash from a passive storage layer into a more active, high-throughput tier in the AI memory hierarchy. The significance of the roadmap is not that HBF will replace HBM4. It will not. Instead, its importance lies in the possibility that HBF can complement HBM4 by absorbing warm data sets that are too large or too costly to keep entirely in HBM but too performance-sensitive to leave in conventional SSD storage.

If Sandisk can bring HBF samples to the market in the HBM4 timeframe, the technology could arrive at a moment when AI datacenter operators urgently need new ways to control memory-driven total cost of ownership. HBM4 will deliver premium bandwidth, but premium memory also carries premium cost, supply complexity, and packaging constraints. HBF could become a capacity amplifier: a tier that allows HBM4 to focus on hot computation while HBF handles model staging, expert weights, embeddings, inactive KV cache pages, adapter libraries, and large retrieval assets.

The roadmap therefore points to a future where AI systems are not designed around a single memory technology. Instead, they will use heterogeneous memory hierarchies. HBM4 will serve the hottest data closest to compute. HBF will provide high-capacity warm storage with improved bandwidth. System DRAM, CXL memory, NVMe SSDs, and object storage will each continue to serve different roles. The winning platforms will be those that combine these tiers with intelligent software for data placement, prefetching, compression, scheduling, and telemetry.

2. Roadmap Context: Why 2027 Matters

Roadmaps matter in datacenter infrastructure because large buyers plan years ahead. AI clusters are not built overnight. A hyperscale AI region, sovereign AI platform, enterprise training cluster, or global inference network requires long procurement cycles, facility planning, power allocation, rack design, cooling strategy, networking architecture, and software qualification. If Sandisk is targeting HBF samples around 2027, the timeline aligns with a broader industry transition toward HBM4-class accelerators and more mature AI-serving architectures.

By 2027, many AI workloads are likely to be more memory-intensive than they are today. Models may use longer context windows by default. Enterprise systems may rely heavily on private retrieval corpora. AI agents may maintain persistent memory and task histories. Multimodal models may process text, image, video, audio, sensor data, and structured enterprise records in a unified pipeline. All of these use cases increase the pressure on memory systems.

The 2027 timeframe is also important because the AI market is expected to move from experimental scaling to cost-controlled scaling. In the early phase of generative AI adoption, many organizations prioritized capability over efficiency. They accepted high inference costs because the strategic value of deployment was uncertain but promising. As usage grows, cost per token, cost per query, cost per image, and cost per agent task become business-critical metrics. Memory hierarchy will be central to improving these metrics.

HBF samples in this timeframe would give OEMs, cloud providers, accelerator vendors, and software companies an opportunity to validate the technology before broad production. Sampling is not the same as mass adoption. It means the ecosystem can start testing bandwidth, latency, endurance, power, thermals, firmware behavior, software interfaces, and workload suitability. If results are strong, HBF could become part of production planning for the late-2020s AI infrastructure cycle.

3. The HBM4 Era and Its Memory Challenge

HBM4 is expected to continue the role of HBM as the premium memory interface for AI accelerators. It should provide higher bandwidth, improved capacity options, and tighter integration with next-generation compute packages. For frontier model training and high-performance inference, this is essential. Transformer workloads, mixture-of-experts routing, multimodal encoders, and large-scale matrix operations all require enormous data movement between memory and compute.

However, each generation of HBM also highlights a structural challenge: premium memory is expensive and difficult to scale. HBM depends on advanced DRAM manufacturing, through-silicon vias, die stacking, complex testing, and advanced packaging. Increasing HBM capacity per accelerator is possible, but it adds cost, power, packaging complexity, and supply-chain risk. HBM4 will likely improve performance, but it will not make memory economics disappear.

The challenge becomes especially visible in inference. Training clusters justify large hardware investments because they create frontier models or strategic capabilities. Inference infrastructure must serve continuous customer traffic, often under strict margin pressure. If every large model deployment requires maximum HBM capacity for all data, inference costs can become difficult to sustain.

The HBM4 era therefore needs complementary memory tiers. HBM4 should be reserved for what it does best: feeding accelerators with hot data at extremely high bandwidth. Other data types should be moved into lower-cost tiers when possible. HBF is relevant precisely because it may provide a better tier for warm data than conventional SSDs, while remaining far more capacity-efficient than HBM.

4. What HBF Is Expected to Contribute

HBF, or High Bandwidth Flash, can be understood as an effort to make NAND flash more useful for performance-sensitive AI infrastructure. Traditional NAND flash is excellent for persistent storage and low cost per bit, but it is not usually treated as an active memory tier. HBF changes the conversation by emphasizing higher bandwidth, greater parallelism, and closer alignment with AI data movement patterns.

The key contribution of HBF is likely to be cost-efficient near-compute capacity. Large AI systems have a growing amount of data that is not hot enough for permanent HBM residency but not cold enough for remote storage. This includes inactive model layers, expert weights in mixture-of-experts systems, embedding indexes for retrieval, document chunks, customer-specific adapter weights, compressed KV cache pages, and checkpoint deltas.

If HBF can provide high sustained read bandwidth, predictable quality of service, and good integration with AI runtimes, it can serve as a staging layer. Data can be moved from HBF into HBM4 before it is needed. With sufficient prefetching and scheduling, some of the latency difference between HBF and HBM4 can be hidden behind computation. This would allow HBM4 capacity to be used more efficiently.

It is important to avoid unrealistic expectations. HBF will not behave like HBM4. NAND has different latency, endurance, and write characteristics. The value of HBF will depend on workloads that can tolerate or hide these differences. In other words, HBF is not a universal memory replacement; it is a specialized tier that becomes powerful when paired with intelligent software.

5. Working Alongside HBM4: A Tiered Architecture

The phrase “work alongside HBM4” should be interpreted as a tiered architecture rather than a replacement strategy. HBM4 sits at the top of the hierarchy, directly connected to the accelerator and responsible for the most bandwidth-sensitive operations. HBF sits below or adjacent to this tier as a high-capacity, high-bandwidth flash layer. System DRAM, CXL-attached memory, NVMe SSDs, and object storage complete the hierarchy.

In a practical AI server, HBM4 may store the active model block, activations, attention data, and hot KV cache. HBF may store less active experts, additional model variants, embeddings, compressed cache pages, and frequently used retrieval assets. When the runtime predicts that certain data will be needed, it promotes the data into HBM4. When data becomes less active, it is evicted or compressed back into HBF.

The efficiency of this architecture depends on predictable access patterns. Transformer inference has some predictable structure because layers execute sequentially. Mixture-of-experts models have more dynamic behavior, but routing information may still enable prefetching. RAG systems can stage retrieval data after query analysis. AI agents can move inactive state out of HBM while keeping the current reasoning path fast.

This hierarchy could be especially important for inference-optimized systems. Training often demands continuous access to huge active tensors, making HBM4 central. Inference workloads, however, frequently contain more exploitable locality. A serving system may know which model, adapter, or retrieval collection a request needs. This allows smarter data placement and creates room for HBF to deliver value.

6. AI Workloads That Could Benefit

HBF is most promising for workloads where capacity pressure is high and data access can be managed intelligently. The following categories are likely to be early candidates for evaluation once samples become available.

6.1 Large-Scale Inference

Large-scale inference is highly sensitive to cost per token. HBM4 will help generate tokens quickly, but HBF can support model staging, cache offload, and multi-model hosting. Service providers that run several model sizes or domain-specific variants may use HBF to keep more assets near compute without filling HBM4 with inactive data.

6.2 Mixture-of-Experts Models

Mixture-of-experts architectures are natural candidates for memory tiering. They contain many expert weights, but only some experts are activated for each token. HBM4 can hold the most active experts, while HBF stores colder experts and promotes them when routing patterns change. The success of this approach depends on routing predictability and prefetch timing.

6.3 Long-Context Applications

Long-context AI systems increase KV cache requirements dramatically. If all sessions keep their cache in HBM4, the system may hit capacity limits quickly. HBF may allow inactive or lower-priority cache pages to be compressed and offloaded. This could support longer conversations, larger batches, or more simultaneous users.

6.4 Retrieval-Augmented Generation

RAG workloads require fast access to large embeddings, vector indexes, and document chunks. HBF can potentially improve the performance of hot retrieval collections by keeping them closer to compute than conventional storage. This is especially relevant for enterprise AI systems that rely on private knowledge bases.

6.5 AI Agents and Persistent Memory

AI agents create state over time. They remember tasks, intermediate plans, tool outputs, files, execution traces, and user preferences. Not all of this state must be in HBM4, but slow access can degrade the agent experience. HBF may provide a middle tier for persistent but frequently revisited agent memory.

7. Supply Chain and Procurement Implications

The introduction of HBF samples around 2027 would have important supply-chain implications. AI datacenter buyers currently face multiple constraints: accelerator availability, HBM supply, advanced packaging capacity, high-power rack deployment, liquid cooling readiness, and networking shortages. Memory is one of the most difficult constraints because it is tightly integrated with accelerator performance.

HBM4 supply will likely be strategically important and potentially tight during early adoption. If every workload requires maximum HBM4 capacity, buyers may face higher costs and longer lead times. HBF offers a possible way to reduce dependence on incremental HBM4 capacity by moving suitable warm data to a lower-cost tier.

For procurement teams, the key question will be whether HBF changes system-level purchasing decisions. If HBF allows a server to host more models, support more tenants, or achieve better utilization with the same number of accelerators, it can influence capital expenditure. If it also reduces model loading delays and improves service density, it can affect operating expenditure.

However, buyers should avoid assuming that sample availability equals immediate production maturity. Samples are for evaluation. Procurement teams should use the sampling phase to test real workloads, validate vendor claims, measure endurance, and assess integration complexity. They should also examine whether HBF will be standardized enough to avoid lock-in and whether multiple platform vendors will support it.

8. Datacenter TCO Impact

The total cost of ownership impact of HBF will depend on how effectively it improves useful work per dollar. AI datacenter TCO includes hardware acquisition, power, cooling, networking, rack space, maintenance, software development, monitoring, and depreciation. A new memory tier improves TCO only if it increases throughput, reduces overprovisioning, lowers latency-related waste, or enables denser deployment without introducing excessive complexity.

The most direct TCO benefit is lower cost per unit of warm capacity. HBM4 will be expensive because it delivers exceptional bandwidth and uses advanced packaging. HBF, based on NAND, should theoretically provide much lower cost per bit. If a large portion of an AI workload’s memory footprint can reside in HBF without harming service quality, the blended memory cost of the system improves.

Another benefit is better accelerator utilization. AI accelerators are costly assets, and idle time is expensive. If HBF can reduce delays when loading model variants, restoring cache state, or accessing retrieval assets, accelerators spend more time generating output. This improves cost per token and cost per query.

HBF may also enable more flexible service tiers. Premium low-latency users may be served with more data resident in HBM4, while standard or batch workloads may rely more heavily on HBF. This allows cloud providers and AI service operators to align infrastructure cost with customer pricing.

The correct benchmark is not theoretical bandwidth. The correct benchmark is economic performance under real traffic: cost per million tokens, tokens per watt, models hosted per rack, p99 latency under peak demand, and recovery time after model switching. HBF’s success will be measured by these system-level outcomes.

9. Software Requirements for Success

HBF will not deliver its full value through hardware alone. AI memory tiering is fundamentally a hardware-software co-design problem. The runtime must know which data belongs in HBM4, which data can remain in HBF, and when data should move. Poor placement decisions can erase the economic advantage of the technology.

Inference engines will need advanced prefetching. For transformer models, layer execution order can guide data movement. For mixture-of-experts models, routing probabilities can inform expert staging. For RAG systems, query planning can identify relevant embedding shards and document chunks. For AI agents, state management can distinguish active reasoning context from background memory.

Cache management will also be critical. KV cache growth is one of the most important memory challenges in long-context inference. A strong HBF-enabled runtime should support cache compression, priority-aware eviction, asynchronous restoration, and latency-sensitive scheduling. It should avoid moving data during the most critical part of a user interaction.

Observability is equally important. Operators need visibility into HBF bandwidth, latency, queue depth, wear status, error correction, temperature, throttling, cache hit rates, and promotion frequency. Without this telemetry, it is impossible to know whether HBF is improving or hurting TCO.

10. Risks, Unknowns, and Adoption Barriers

The first major risk is performance uncertainty. Until samples are tested under real AI workloads, it is difficult to know how HBF will behave in production-like environments. Sequential bandwidth may look promising, but AI workloads can include irregular access patterns, mixed reads and writes, and latency-sensitive bursts.

The second risk is software readiness. A new memory tier requires support across drivers, operating systems, runtime libraries, inference engines, orchestration systems, and monitoring platforms. If this software support arrives slowly, HBF adoption may be delayed even if the hardware is promising.

The third risk is endurance. NAND flash has finite write endurance. Workloads that constantly offload and restore cache pages or frequently update retrieval indexes may create write amplification. HBF systems must be designed with endurance-aware policies, and buyers should demand transparent lifetime models.

The fourth risk is ecosystem fragmentation. If HBF implementations are proprietary or incompatible across vendors, software developers may hesitate to optimize for them. Standardization will be essential for broad adoption.

The fifth risk is timing. HBM4 platforms will have their own development and qualification schedules. For HBF to work alongside HBM4, alignment across accelerator vendors, server OEMs, memory suppliers, and software providers is necessary. Any mismatch could push adoption later than the sample timeline suggests.

11. Market Outlook Toward 2027 and Beyond

Looking toward 2027 and beyond, the AI memory market is likely to become more segmented. HBM4 will anchor premium training and high-performance inference platforms. DRAM and CXL memory expansion will provide flexible host-side capacity. HBF and other high-bandwidth flash technologies may compete for the warm memory tier. Conventional SSDs and object storage will continue to support massive persistent data stores.

The strongest market pull for HBF will come from inference economics. As AI applications reach more users, inference volume may become the dominant cost driver. Companies will need to serve larger models and longer contexts without allowing infrastructure cost to grow linearly. HBF may help decouple model capability from HBM-only capacity expansion.

Another driver will be the rise of multi-tenant AI platforms. Cloud providers and enterprise AI platforms increasingly need to host many customers, adapters, private knowledge bases, and security policies on shared infrastructure. HBF can support a larger pool of near-compute assets, improving service flexibility.

However, market success will depend on credible proof points. Buyers will look for benchmark data, reference architectures, ecosystem partnerships, software support, and clear migration paths. If HBF demonstrates measurable reductions in cost per token and improvements in model density, adoption could accelerate. If the technology remains difficult to integrate or offers only marginal benefits, it may stay limited to niche deployments.

12. Conclusion

Sandisk’s HBF roadmap targeting 2027 samples alongside HBM4 represents a potentially important step in the evolution of AI memory architecture. HBM4 will continue to provide the highest-performance memory tier for next-generation AI accelerators, but it will not solve every memory challenge by itself. AI datacenters increasingly need more capacity, more flexible data placement, and lower total cost of ownership.

HBF’s opportunity is to become a practical warm tier between HBM4 and conventional storage. If it can deliver high sustained bandwidth, acceptable latency, strong endurance, and standardized integration, it may help AI operators use HBM4 more efficiently. Rather than filling premium HBM4 with every model asset, operators could reserve HBM4 for active computation and rely on HBF for staging and warm data management.

The roadmap should be interpreted carefully. Samples in 2027 would mark the beginning of ecosystem validation, not immediate universal deployment. The technology must prove itself through real workloads, strong software support, and clear economic value. But if those pieces come together, HBF could play a meaningful role in the late-2020s AI infrastructure stack.

The broader lesson is that AI performance is no longer only about faster accelerators. It is about moving the right data to the right memory tier at the right time. In that future, HBM4 and HBF are not competitors. They are complementary tools for building scalable, cost-efficient, and memory-aware AI datacenters.

FAQ