Nvidia's core GPU business remains dominant, but its ambitious new bet on specialized decode hardware is facing unexpected pressure from free, open-source software innovations that solve the same problem without requiring customers to buy additional racks. The company just posted its biggest quarter ever, with $81.6 billion in revenue and $75.2 billion from data center operations, but investors are now questioning whether a second hardware purchase will actually happen at scale.

What Is Nvidia's Groq 3 LPX, and Why Does It Matter?

Nvidia is launching a specialized piece of hardware called the Groq 3 LPX, designed to handle a specific bottleneck in artificial intelligence inference, the process where trained models generate responses to user queries. The LPX contains 256 specialized processors, each with 500 megabytes of ultra-fast memory running at 150 terabytes per second of bandwidth, roughly seven times faster than standard GPU memory. When paired with Nvidia's Vera Rubin GPU system, the company claims it delivers up to 35 times higher inference throughput per megawatt for trillion-parameter models.

The pitch sounds compelling: for companies running the most demanding artificial intelligence workloads, this specialized hardware promises dramatically faster performance. But here is the catch. Customers who have already invested in Vera Rubin GPUs must make a separate purchase decision to add LPX to their data centers. That second check only gets written if the performance improvement justifies the additional cost.

How Does AI Inference Actually Work, and Why Is Decode the Bottleneck?

To understand why Nvidia is betting on specialized decode hardware, it helps to know how large language models actually generate responses. The process splits into two distinct phases. First, prefill processes the input prompt and generates the initial memory state. Second, decode generates output tokens one at a time, using that memory state while handling sustained pressure from active users, long outputs, and large context windows.

Decode is slower, more memory-intensive, and harder to scale efficiently. The memory structure at the center of this pressure is called the KV cache, which grows with context length and must be read repeatedly for every generated token. For long-context artificial intelligence workloads, the KV cache can consume the majority of available GPU memory. This is precisely the bottleneck that LPX is designed to solve.

Why Is DeepSeek's DSpark Threatening Nvidia's Hardware Strategy?

On June 27, 2026, Chinese artificial intelligence lab DeepSeek released DSpark, a speculative decoding module that attacks the decode bottleneck using pure software innovation. The mechanism works by having a smaller draft model propose multiple tokens at once, while the large target model verifies them in parallel. When the draft is correct, multiple tokens are accepted in a single step, reducing the number of full decode passes required per output and lowering the memory and compute burden per token.

DeepSeek reports per-user generation speed improving 60 percent to 85 percent on its V4-Flash model and 57 percent to 78 percent on V4-Pro, with throughput at a fixed service level improving 51 percent. More importantly, DSpark is open-sourced under the MIT license, meaning anyone can use it for free. The companion DeepSpec training framework already extends to Qwen and Gemma model families, so the efficiency gains are spreading beyond DeepSeek's own ecosystem.

How to Evaluate Nvidia's LPX Attach Rate as an Investor

• Separate Purchase Decision Risk: LPX requires customers to write a second check after already committing to Vera Rubin GPUs, which only happens if the performance gap justifies the additional cost and complexity.
• Software Solutions Reducing Hardware Urgency: DeepSeek's MLA architecture, carried through every model generation since V2, stores compressed context instead of full memory state, reducing memory requirements by roughly 90 percent for million-token conversations, which decreases the urgency for hardware designed to handle memory pressure.
• Competing Industry Solutions: AWS and Cerebras announced a multiyear collaboration in March 2026 pairing Trainium 3 for prefill with Cerebras CS-3 for decode, launching through Amazon Bedrock in the second half of 2026 on the same timeline as LPX, giving customers a hyperscaler-native alternative.
• Open-Source Momentum: DSpark and MLA are free, MIT-licensed innovations already in production and spreading across model families, creating a pattern where software solves the decode problem before hardware ever sees it.

The causal chain for investors is straightforward: decode is a memory and latency problem; LPX is a hardware solution to that problem; DSpark and MLA are software solutions to the same problem. They are open, free, and already in production.

Is Nvidia's LPX Strategy Doomed, or Just Uncertain?

Nvidia's own LPX architecture actually supports speculative decoding, the same technique DSpark uses. Dynamo is designed to orchestrate draft-and-verify workflows across the GPU-LPU combination. So DSpark and LPX are not simply in opposition. The more pointed concern is that DSpark running on general Rubin GPUs alone, without LPX attached, delivers enough inference efficiency that the second rack becomes optional for most workloads.

The broader context matters here. The case for separating prefill and decode onto specialized hardware is not Nvidia's alone. It is the conclusion the entire industry has reached simultaneously, which validates the underlying thesis but complicates the investment case for LPX specifically. If every major cloud provider reaches the same architectural conclusion and builds their own answer to it, Nvidia's LPX attach rate becomes a question of whether customers who buy Rubin GPUs also buy LPX as a second rack, or whether they route their most latency-sensitive decode workloads to a hyperscaler-native alternative instead.

The market noticed DeepSeek's speed numbers and moved on. What those speed numbers actually signal is more important than the benchmark itself. A Chinese artificial intelligence lab keeps finding ways to make inference faster using software and open weights, at no cost to anyone who wants to use it. Meanwhile, Nvidia is ramping a specialized decode rack that requires a separate purchase decision on top of the GPU platform customers already depend on. The question the market is not asking is whether that second check gets written at scale, or whether DSpark and the architectural innovations underneath it are quietly making the answer no.