The next bottleneck in spatial biology may not be biology, but memory.
You check the size of the spatial transcriptomics data files and provision sufficient disk space. Once the analysis begins, however, it takes longer than expected and eventually crashes due to insufficient memory. Adding GPUs does little to improve the speed.
This happens repeatedly because the data challenge in spatial transcriptomics is not simply a matter of storage capacity. Actual performance is determined by a combination of RAM and VRAM, storage I/O speed, CPU preprocessing, network data transfers, and the intermediate data generated during analysis.
File size is visible. But the peak memory required while an analysis is running is much harder to see.
As Resolution Increases, Matrices Grow Even Faster
Spatial transcriptomics data generally represents the relationships between locations or cells and genes as massive matrices. As measurement resolution increases and more transcripts are detected, not only the number of rows and columns but also the number of entries containing actual values, known as non-zero entries, rises dramatically.
Raw data can be compressed as a sparse matrix, in which most values are empty. During filtering, normalization, integration, and cell-level operations, however, data may be copied and new intermediate results generated. If a particular operation does not make sufficient use of the sparse structure, it may require several times more memory than the size of the stored file.
The more important question, therefore, is not “How many terabytes can we store?” but “At which stage of the analysis does the data expand, and how?”
GPUs Do Not Solve Every Bottleneck
Many spatial transcriptomics operations are well suited to GPUs, but not every stage of the analysis is GPU-centric. Reading and organizing large files, merging metadata, filtering sparse matrices, and converting formats can depend heavily on CPU performance, memory bandwidth, and I/O performance.
Even if GPU computation itself is fast, the overall analysis will not accelerate if transferring data from storage to RAM and then to GPU VRAM takes too long. The result is an expensive accelerator left waiting for data.
Intuitively, purchasing more computing resources may seem like the solution. In practice, however, data movement may be slower than computation, or the peak memory requirement of a single stage may force the entire system to be dramatically overprovisioned. Designing spatial transcriptomics infrastructure is not about choosing between CPUs and GPUs. It is about achieving the right balance among compute, memory, and storage.
There Is No Single Right Answer Between Cloud and On-Premises
The cloud makes it easier to provision resources elastically and enables collaboration among researchers across multiple regions. On the other hand, repeatedly transferring and storing large datasets over long periods can lead to accumulating costs. Security and data sovereignty must also be considered when working with patient-derived data.
On-premises environments offer advantages in data control and repeated large-scale analyses, but they require upfront investment and operational staff. Equipment provisioned for peak demand may also remain idle during normal periods. For small cohorts or one-off projects, existing servers or cloud instances may be more economical.
What matters is not adhering to a particular approach, but comparing total costs based on analysis frequency, data transfer volume, security requirements, and processing time.
The Hidden Costs Portrai Has Observed
In a specific CPU preprocessing proof of concept conducted by Portrai, peak memory usage reached approximately 2.18 TiB while processing around 34.78 million cells and 29.1 billion non-zero expression entries. This figure does not apply uniformly to all spatial transcriptomics analyses. It was observed with the specific data structure and filter_empty workload used in that test.
The key lesson from this experience was that a particular preprocessing stage, rather than the overall data size, determined the memory specifications of the entire system. In some cases, optimizing the analysis code may be sufficient. If datasets repeatedly exceed this scale, however, the memory hierarchy must be redesigned alongside software improvements.
A Tiered Architecture Instead of a Single Expensive Server
In the short term, data should be kept in sparse, chunked formats, with out-of-core computing and caching used to load only the portions required. Simply reducing unnecessary copying and format conversions can significantly lower costs.
In the medium term, fast DRAM can be tiered with relatively affordable enterprise SSDs, creating an architecture that does not require all data to remain resident in expensive memory. The way data is generated should also be co-designed with the filtering, integration, and visualization processes that follow.
In the long term, potential alternatives include CXL-based memory expansion, DPUs that share the burden of data movement, and NPUs suited to specific operations. Whether these technologies can improve both performance and cost for real spatial transcriptomics workloads must be validated through benchmarks. New hardware may also introduce additional software complexity and operating costs. Under conventional server architectures, however, certain tasks currently require disproportionately large infrastructure, and this is likely to remain one of the most important bottlenecks for the foreseeable future.
The Next Five Years
In an optimistic scenario, data formats, memory hierarchies, and accelerators optimized for spatial biology will advance together, making analyses of tens of millions of cells part of standard workflows. More realistically, hybrid architectures combining cloud and on-premises environments, as well as general-purpose hardware and next-generation infrastructure, are likely to become more widespread for different workloads.
Conversely, if the race for throughput continues without reducing the total cost per analysis, the gap may widen between institutions that can generate data and those that can actually use it. This is already happening. Institutions purchase expensive spatial transcriptomics equipment and produce data, yet more and more of that data accumulates without blossoming into publications or research that makes the world a better place.
The next bottleneck in spatial biology may not be biology. It may be memory architecture.
Yet even if large-scale data can be processed quickly, discovering meaning within the results is an entirely different challenge. In the next article, we will explore the democratization of the analysis experience: enabling researchers to move beyond the burden of coding and server management and focus on the scientific questions that truly matter.
