Samsung’s memory meltdown — and the surprising 2025 comeback
The Samsung Semiconductor Memory story is one of the most remarkable in the technology sector. From its humble beginnings in the 1990s—when a faulty rice cooker caused a customer to question the quality of Samsung’s memory—the company rose to dominate an industry once led by the United States.
As most American memory companies disappeared, Samsung surged ahead and began to rival Intel as the most influential semiconductor company in the world. With more than 40% market share for over a decade, Samsung appeared untouchable, and major long-term plans were already in motion to reinforce its leadership.
In early 2021, Samsung began developing PIM (Processing in Memory), building on its HBM (High Bandwidth Memory) foundation. A year later, it demonstrated the technology running on an AMD GPU. But despite the promising start, Samsung soon hit turbulence.
A series of qualification issues with HBM3E—particularly versus SK Hynix—exposed deeper operational problems. By 2024, Samsung’s HBM market share fell from over 60% to under 20%, while SK Hynix became Nvidia’s preferred supplier. It was a dramatic shift for a company once considered unbeatable.
Yet, despite this “horror year,” recovery signs have emerged. Samsung’s Q3-25 results were its strongest in three years, approaching previous peak levels. Revenue rebounded sharply, and net profit more than doubled year-on-year. As always, the memory division sits at the heart of this performance.
Although Samsung is a complex conglomerate—with mobile phones, displays, consumer electronics, and components—the memory business still dictates the company’s profitability.
When memory is booming, Samsung becomes a profit machine; when it falls, the entire company feels it. In the last upcycle, memory alone delivered more operating profit than all other divisions combined.
The latest surge in memory operating profit is not primarily due to increased output. Samsung has been running at full capacity for years—sometimes profitably, sometimes at heavy losses. Instead, most of the recent revenue uplift is driven by steep price increases across the global memory market.
Inventory behaviour confirms this. Finished goods have declined, but not nearly enough to account for the revenue spike. Work-in-progress inventory has flatlined, suggesting production levels are not rising meaningfully. In short: demand is surging, prices are rising, but Samsung’s ability to increase supply is constrained.
A major reason is underinvestment. Like other memory manufacturers, Samsung was shaken by the severity of the last downturn and pulled back hard on CapEx. Recent investment levels have been far below what is required even to maintain capacity, let alone expand it. In Q3, Samsung should have spent around $9.6B just on maintenance CapEx, yet invested only $5.6B.
This raises an important question: what happens when memory demand continues to grow, but supply cannot keep up?
To explore the full analysis—including charts, market dynamics, and what this means for 2025—click the link to read the complete story.
Claus is offering readers of The IC Resource 40% off access to his exclusive insights.
Claus is an expert at peeling back the layers of polished corporate messaging. His curiosity drives him to explore semiconductor companies, focusing on the complexities of their supply chains—both upstream and downstream. He compares companies to their competitors, uncovering key insights that often go unnoticed.
Written by Claus Aasholm
Subscribe to Claus’ Substack with a 40% discount
Big photonics powering precision
Written by Stratos Kehayas
President, G&H Photonics Division
Photonics is rapidly becoming a foundational technology in the life sciences, driving advances in how biological systems are visualized, measured, and interpreted. This article emphasizes that improvements in photonic components and systems – greater spectral precision, better stability, and more efficient light delivery – are enabling scientific capabilities that were previously out of reach. These developments are especially transformative in the following areas: biological imaging, quantum sensing, and AI-enabled analysis.
Quantum sensing: extending detection beyond classical limits
The article positions quantum sensing as an emerging and potentially revolutionary photonic frontier. By leveraging tools such as fiber-coupled acousto-optic modulators and high-precision collimators—components essential for manipulating atoms with lasers – quantum sensors can achieve sensitivities far beyond those of classical optical instruments.
While quantum sensing is already established in fields like navigation and timing, its translation to biological applications is still underway. If successfully adapted, quantum sensors could enable measurement of biological signals at unprecedented scales: detecting minute biomolecular changes, mapping neural activity with extreme sensitivity, or characterizing tissue based on subtle quantum-level interactions. Though still early, this technology could open an entirely new class of life-science instrumentation.
AI: extracting meaning from complex photonic data
As photonic instruments produce increasingly rich datasets, AI becomes essential for interpretation. Machine-learning models can analyze subtle patterns in optical images, detect early indications of disease, or interpret spectroscopic signatures beyond what human observers can discern.
AI also enhances the imaging process itself – through noise reduction, adaptive optics control, and automated segmentation of biological structures.
Together, photonics and AI form a tightly integrated system: photonics generates precise, high-quality data, and AI transforms it into actionable biological or clinical insights.
Imaging: richer and more precise biological information
In imaging, the evolution of photonics is allowing researchers to extract deeper and more comprehensive insights from biological samples. Modern fluorescence microscopy and related imaging methods now rely on highly precise photonic control: rapid wavelength switching, fine spectral filtering, and stable light delivery. These capabilities make multiparametric imaging possible – where multiple fluorescent tags or biomarkers can be imaged simultaneously within a single sample.
Such high-content imaging extends far beyond structural visualization. Researchers can track dynamic processes, measure molecular interactions, and observe multiple cellular pathways at once. Improvements in photonic hardware, including low-noise light sources and advanced optical filters, ensure that the data collected is accurate and reproducible. As a result, imaging is poised to play an expanding role in drug discovery, disease modeling, and personalized diagnostics.
The emerging ecosystem
The future of life-science technology lies in systems that integrate all three elements – advanced imaging, quantum-level measurement, and intelligent AI-driven analysis – creating a new generation of precise, high-impact tools for both research and medicine.”
Transformers At The Edge: Efficient LLM Deployment
Since the groundbreaking 2017 publication of “Attention Is All You Need,” the transformer architecture has fundamentally reshaped artificial intelligence research and development. This innovation laid the foundation for Large Language Models (LLMs) and Video Language Models (VLMs), fueling a wave of productization across the industry. A defining milestone was the public launch of ChatGPT in November 2022, which brought transformer-powered AI into mainstream use. Since then, LLMs have enabled a broad spectrum of applications, from conversational agents to advancements in medical research.
However, running these LLMs efficiently presents substantial challenges, particularly on edge computing devices and legacy hardware architectures that were designed before the widespread adoption of large language models.
One of the significant difficulties facing AI processors is the sheer size of LLMs compared to prior state-of-the-art CNNs (Convolutional Neural Networks), RNNs (Recurrent Neural Networks), and other network types.
Among these CNNs and RNNs, an 85 million-parameter model would have been considered large. In comparison, even a modestly sized LLM might have 1B parameters, while models with 8B parameters, and larger are commonplace. Said plainly, there is no mass market, cost-effective method to load that many parameters on a single chip; thus, pre-existing solutions may not be effective.
Consider Llama 3.2, Meta’s latest generation of LLMs, which introduced significant advancements in both text and multimodal (text + vision) AI capabilities. This release expands on previous versions with new model variants and features designed for both enterprise and edge-device deployment. Llama 3.2 contains 1 billion parameters. Furthermore, attention operations (O) increase with context size (n).
During prefill, where operations are predominantly done in parallel, the compute load is a function of the square of the context size, O(n2).
The prefill phase is compute-bound, meaning its speed is limited by the raw computational power of the hardware. In contrast, decode is predominantly sequential, and the compute load is an order of magnitude smaller, O(n)*n per token. However, the decode phase is dominated by memory access speed rather than compute power, and the per-token compute cost can be orders of magnitude higher than during prefill.
Exploring LLM Inference Flow
The LLM inference flow begins with a user prompt, which is a sentence spoken or input by the user. The user prompt is first translated into what is called a token, a mathematical representation of the user prompt, using a processing mechanism aptly named the “tokenizer.” The token is then sent to the inference processing steps, where they are divided into two phases: the prefill phase and the decode phase.
In the prefill phase, all the tokens are sent at once through a series of transformer blocks. At each transformer layer, the Key (K) and Value (V) vectors for each token are stored in a cache — this is called the KV cache or attention cache. This cache is then used to make subsequent generations faster. In the decode phase, one response at a time is generated, as this part is sequential in nature.
During the prefill stage, the model needs to compute attention over all previous tokens. However, during the decode stage, the processor only needs to compute attention over the new token because it reuses the cached values from prefill. This enables efficient autoregressive decoding — the processor doesn’t need to recompute everything from scratch each time. The prefill stage is compute-heavy because it processes the entire input prompt.
Compute cost scales linearly with prompt length – if the prompt is N tokens and the model has L layers, the processor must do N × L full forward passes. However, the decode stage is relatively fast after that because the system processes only one token at a time, reusing the cache. This raises an issue, though – as the phases are very different in compute, memory, and power, how can a solution be enabled which is optimal for both?
LLM inference
Let’s also examine the runtime changes between traditional AI networks and LLMs. Traditional CNNs have a simple, monolithic runtime with only two phases: Data loading phase and Inference phase. LLMs introduce a multi-phase runtime system with five distinct phases, each with different computational and memory requirements.
Prefill Phase
– Processes the initial user prompt by embedding and tokenizing the entire input sequence.
– Runs all transformer layers in full sequence mode (dense computation)
– Initializes and populates the Key-Value (KV) cache with attention values.
– Generally has higher per-token latency due to sequence processing.
– May use microbatching for lengthy inputs.
Decode Phase
– Generates output tokens one at a time autoregressively.
– Only processes the last generated token per step.
– Retrieves past tokens from the KV cache for efficiency.
– Computes self-attention only against past tokens.
– Highly optimized with KV caching and batching.
Inactive Phase
– No computation occurs, but the sequence remains “alive” in memory.
– Occurs when waiting for new user input in streaming/chat interfaces.
– KV cache remains in memory (costly resource usage).
– Can become a bottleneck in high-throughput systems with many cached sequences.
Follow-up Prefill
– Triggered when new user input is appended to a partially generated sequence (multi-turn conversations).
– Processes new input as a short prefill segment, appending to existing cached content.
– Updates KV cache with new tokens.
– Distinct from initial prefill as it operates on shorter, appended segments.
Retired Phase
– Sequence is terminated and removed from the active pool
– KV cache is freed and resources are released.
– Triggered by conversation completion, user cancellation, or timeout.
– Frees memory and scheduling capacity for other sequences
This multi-phase complexity significantly exacerbates deployment difficulties for LLMs compared to traditional AI networks. Managing multiple phases simultaneously, maintaining expensive KV caches, and handling dynamic transitions between phases create substantial challenges for efficient LLM deployment and resource management.
New Architectures for LLMs
Large Language Models introduce distinct challenges for inference hardware and software systems, including complex model architectures, demanding runtime computations, transformer-specific operations, and implementation considerations. To address these requirements, AI processing platforms must advance to accommodate diverse data representations, support multiple precision formats, deliver enhanced computational throughput, and enable efficient multi-core processing architectures.
Expedera explores this further in their white paper at https://www.expedera.com/transformers-at-the-edge/
Written by Paul Karazuba
VP Marketing, Expedera AI
Shaping the Future of UK Deep Tech and Semiconductors
Written by;
James Cunningham – Business Development Director
As well as writing long-term success stories, our aim is to contribute to the health of the industry as a whole.
IC Resources has been supporting the deep-tech community since 1999. We are an independent, founder-led recruitment consultancy with a team of 70 specialist consultants who operate across Semiconductor, Electronics, Software, AI, Photonics and Quantum.
Our work spans technical, operations, sales and marketing and executive search, and we offer both global reach and local expertise with offices in Reading, London, Munich, Austin and Pennsylvania.
2025 Highlights
As most will know, 2025 has been clearly a year of two markets; AI & ‘The Rest’. Concern and uncertainty over the impact of Tariffs has caused confusion and caution – but in most cases the question remains – outside of AI, is the underlying demand actually that high right now?
Put simply; investment, growth and hiring has been highly centred around the massive rise of AI and the ecosystem surrounding it. Significant strategic acquisitions came as a result of companies wanting to rapidly implement proven AI solutions – Qualcomm bought Alphawave, Softbank – Ampere. But funding available outside of AI has been quite limited which has presented some very real growth challenges across the industry.
Sectors that I believe are worth mentioning specifically:
Automotive
The western automotive industry is changing as traditional vehicle manufacturers now have to pull together major componentry elements from multiple, independent OEMs or Tier 1/2/3 sources and ensure that the software still communicates effectively and securely.
They face immediate competition from the east where, since Covid, software defined vehicles have been designed from the ground up, do not face these issues and are hitting the market at a fraction of the cost. This is creating an unprecedented level of challenge to a once dominant vertical.
Robotics
Interesting new developments in the field of Robotics, following the integration of AI – is iRobot is becoming a reality? Definitely a growth space to watch.
Battery life
Some really interesting innovations in the ultra low power space – from Edge AI processors to IOT sensors and wireless communication solutions that are breaking barriers in power consumption. Even solutions that can charge micro batteries from ambient energy alone – thus creating an ‘infinite energy source’. With the right funding connectivity is set to increase.
Fabs
Companies committed massive investments in R&D and new fabrication plants to meet demand, with global capital expenditures expected to reach around $185 billion in 2025 and at least 18 new fabs began construction worldwide. But there’s a new player on the block… The leading nodes are ever-advancing but there has been one clear leader for some time now in TSMC. Of late though an interesting disruptor is making moves to take on the leader head on at the most advanced nodes – why play catch up when you can join right at the finish line?
Space comms
The sudden ability to put satellites in space faster and more cost effectively than ever before has driven huge innovation in the uplink and downlink communications. We expect further innovation to come, taking advantage of this fast emerging market.
As a wrap up – we wish everyone a fantastic festive season and are looking forward to the next year of unforeseen market challenges. Personally I am keen to see new companies emerge victorious and whether AI is in fact the ‘bubble’ that people seem to suspect. Personally I think that once we get to see beyond the ‘search engine / creative writing capabilities the bubble may in fact just keep growing. Here’s to 2026!
For more information on any of the articles featured in this edition of The IC Resource, or, if you would like to contribute to our future editions, please get in touch – james.cunningham@ic-resources.com or +44 (0)118 988 1166.
1530 Arlington Business
Park, Berkshire.
RG7 4SA
© Intellectual Capital Resources Ltd 2024
