Hardware Is Eating AI,
and AI Is Eating Everything Else

When Software Was Free and Scale Was Infinite

In 2011, Marc Andreessen published his famous essay arguing that software was eating the world. The thesis was simple: software has near-zero marginal cost. Once you write a piece of code, distributing it to a million users costs almost nothing. That economic reality gave software companies a structural advantage over every traditional industry. Retail, media, finance, transportation, healthcare. One by one, they fell.

And for fifteen years, the thesis held. Netflix displaced Blockbuster. Spotify reshaped music. Uber redrew transportation. Amazon swallowed retail. The common thread was that software's weightlessness, its freedom from atoms, was its superpower.

But something fundamental has changed. The next wave of technology, artificial intelligence, is not weightless. It is brutally, inescapably physical. Every model trained, every inference served, every reasoning chain computed demands GPUs, high-bandwidth memory, advanced packaging, electrical power, cooling infrastructure, and increasingly, water. The era of zero-marginal-cost disruption is over. Welcome to the era where atoms are eating bits.

Hardware Is Eating AI

If I had to write the 2026 version of Andreessen's essay, the headline would be: "Hardware is eating AI, and AI is eating everything else."

The logic is simple. AI is doing exactly what software did, disrupting every industry it touches. But unlike software, AI cannot scale for free. Its supply chain has chokepoints that look more like the oil industry than the software industry. The companies that control scarce hardware hold enormous power, almost like OPEC did with crude.

Look at the bottleneck stack. It is not just GPUs. The binding constraint is actually memory and packaging:

All of this creates a cascading supply crisis. NVIDIA has reportedly cut consumer RTX 5000-series GPU production by 30 to 40% to prioritize data center chips. Micron discontinued its entire consumer Crucial memory lineup to focus on AI. PC prices are climbing. Even game developers are being forced into unplanned optimization work because RAM costs have restructured what is affordable to ship.

This is not a GPU shortage. It is a memory shortage, a packaging shortage, a power shortage, and a supply chain architecture problem, all compounding at once.

What Do the Projections Actually Say?

The question everyone is asking: is this a structural shift or a cyclical bottleneck? The data suggests it is cyclical, but with a longer tail than most people expect.

Supply-side timeline

DRAM consumption by AI: the elephant in the room

One number tells the whole story: AI data centers are projected to consume approximately 70% of all global DRAM production in 2026. That single statistic explains why everything from laptops to gaming consoles to Raspberry Pis is getting more expensive. The memory market has been structurally redirected.

TrendForce researchers have described this reallocation of memory capacity toward AI as "permanent." Even when new capacity comes online, a significant share will remain tied to AI customers. This is not a temporary diversion. It is a structural rebalancing of who gets to buy memory and at what price.

Five Forces That Will Break the Bottleneck

If this were purely about physics, finite silicon, finite fabs, the shortage would be permanent. But technology markets correct themselves: when something is expensive enough, the entire industry innovates around it. The relief comes from five directions.

1. Alternative accelerators and inference-specific silicon

The GPU monoculture is ending. Google's TPU v6e already delivers up to four times better performance per dollar for inference compared to NVIDIA's H100. Amazon's Trainium, Intel's Gaudi, and various FPGA-based solutions are diversifying the compute supply chain. XPU spending is projected to grow 22% in 2026, and by 2027, these alternatives may account for a larger share of AI hardware budgets than GPUs.

But the most telling signal came on Christmas Eve 2025, when NVIDIA itself paid $20 billion to license Groq's Language Processing Unit (LPU) technology. Think about what that means: the company that dominates GPUs spent its largest deal ever to acquire an architecture that does not use GPUs at all.

Groq's LPU completely bypasses the HBM memory bottleneck that sits at the center of the 2026 shortage. Instead of using off-chip HBM (the scarce, expensive memory that SK Hynix, Samsung, and Micron cannot produce fast enough), Groq built its chips with on-chip SRAM as primary weight storage. SRAM is roughly 100 times faster than HBM and 20 times more energy-efficient per bit of data access. The trade-off is capacity: each LPU chip holds only about 230 to 500 MB, so you need hundreds of chips linked together to run a large model. But those chips do not consume any HBM, do not require TSMC's constrained CoWoS packaging, can run on older 14nm fabrication nodes, and are air-cooled with no liquid cooling infrastructure needed.

The inference efficiency numbers make the strategic logic clear. NVIDIA's own published figures for the Groq 3 LPU show approximately 150 tokens per watt, compared to roughly 4.3 tokens per watt for the H100 in comparable model serving. That is a 35x improvement in inference throughput per megawatt. For data center operators running millions of inference requests daily, that changes the economics entirely.

This matters for the shortage because AI is shifting from a training-dominated phase to an inference-dominated phase. Deloitte estimates inference workloads now account for roughly two-thirds of all AI compute, up from one-third in 2023. The market for inference-optimized chips alone is projected to exceed $50 billion in 2026. If a meaningful share of those inference workloads can move to architectures that do not consume HBM, CoWoS slots, or cutting-edge fab capacity, it directly relieves pressure on the entire supply chain.

The point extends beyond Groq. Arm launched its AGI CPU in March 2026, designed specifically for the decode phase of AI inference and agentic AI orchestration. Arm's CEO predicted the chip alone would generate $15 billion in revenue by 2031, and forecast that emerging inference workloads would quadruple demand for CPUs. Intel and SambaNova announced a multi-year collaboration built on a modular architecture where GPUs handle prefill, SambaNova's RDUs handle decode, and Xeon CPUs orchestrate the system. Microsoft shipped Maia 200, its own 3nm inference accelerator, already deployed in Azure powering Copilot and GPT workloads. Cerebras continues to push wafer-scale SRAM-based inference at speeds six times faster than Groq on comparable models.

Not every AI workload needs a top-tier GPU with HBM. Many production inference tasks, particularly in enterprise settings, can run efficiently on CPUs, specialized ASICs, or lighter accelerators. The industry is waking up to the fact that the right chip for the job is no longer always a GPU, and that realization is itself a release valve on the shortage.

2. Model efficiency is improving fast

The era of "just throw more GPUs at it" is giving way to smarter engineering. Smaller, distilled models that run on a fraction of the hardware are becoming viable for most production workloads. Techniques like quantization, mixture-of-experts, and speculative decoding are reducing the compute cost per inference by orders of magnitude.

3. Processing-in-memory will change the game

The real bottleneck in 2026 AI is not computation. It is memory bandwidth. The cost of shuttling data between memory and processor is where most energy and time is consumed. Samsung and SK Hynix are investing heavily in Processing-in-Memory (PIM) architectures, where computation happens directly inside the memory chips. This could significantly reduce the data movement problem that drives current hardware costs.

4. New fabs are coming online

The capital is flowing. Intel, TSMC, Samsung, and Micron all have major fab construction projects underway. The challenge is physics: chip factories take 3 to 5 years to build, equip, and certify. But by 2028, the supply side will look very different from today.

5. Neuromorphic and novel architectures

The human brain runs on about 20 watts of power. Current AI clusters use megawatts. That gap represents one of the largest efficiency opportunities in the history of computing. Research into neuromorphic chips, processors that mimic the brain's use of spikes and pulses rather than continuous electricity, is accelerating because the current power economics are unsustainable.

Custom Silicon and Software-Defined Efficiency: Broadcom's Two-Pronged Answer

One company is uniquely positioned to attack the hardware crisis from both sides of the equation: Broadcom. On one hand, it is building the custom silicon that breaks the GPU monoculture. On the other, its VMware Cloud Foundation platform helps enterprises squeeze maximum value from the hardware they already own. That combination makes Broadcom one of the most important players in the AI infrastructure story right now.

The TPU partnership: custom silicon at scale

On April 6, 2026, Broadcom filed an 8-K disclosing a long-term agreement with Google to develop and supply future generations of custom Tensor Processing Units through 2031. This is not a chip order. It is a five-year infrastructure commitment covering processors, networking fabric, and the entire AI rack architecture.

The deal also pulled in Anthropic. Under the agreement, Anthropic will access approximately 3.5 gigawatts of next-generation TPU-based compute capacity starting in 2027. To put that in perspective, 3.5 gigawatts rivals the energy output of multiple nuclear power plants. The build-out is estimated to cost between $120 billion and $175 billion, making it one of the largest infrastructure investments in technological history.

Why does this matter for the hardware crisis? Because custom TPUs deliver better performance per watt than general-purpose GPUs for specific workloads. Google's latest Ironwood TPU v7, built on a 3-nanometer process, offers a four-fold performance improvement over previous generations while consuming less power per teraflop. When you can do more work per chip, you need fewer chips. When you need fewer chips, you ease the pressure on memory, packaging, and power.

Broadcom's AI revenue trajectory tells the story. From $4.4 billion in Q2 FY2025 to $8.4 billion in Q1 FY2026, that is 106% year-over-year growth. The company is targeting $100 billion in AI chip revenue by 2027, with a $73 billion backlog already in hand. Broadcom has captured 60 to 70% of the custom AI accelerator market, and its ability to integrate custom silicon with high-speed networking (Tomahawk and Jericho Ethernet switches, industry-leading SerDes IP) creates a full-stack offering that competitors cannot easily replicate.

Anthropic's approach here is worth noting. The company is explicitly taking a multi-hardware strategy, using AWS Trainium, Google TPUs, and NVIDIA GPUs in parallel. The market is not consolidating around one chip. It is fragmenting into strategic, workload-specific choices. That fragmentation is itself a release valve on the shortage.

VCF 9.0: do more with the hardware you already have

The other half of Broadcom's answer comes from VMware Cloud Foundation. While the TPU partnership addresses the supply side (build better chips), VCF attacks the demand side (use existing hardware more efficiently).

Broadcom's VCF division has been direct about this. Krish Prasad, SVP and GM of the VMware Cloud Foundation Division, said it plainly: the answer to a hardware crisis is not more hardware. It is smarter software. VCF 9.0 was engineered with the 2026 supply crunch in mind, built around four pillars of efficiency.

NVMe memory tiering deserves special attention because it directly addresses the memory crisis at the heart of the 2026 shortage. The way it works: ESXi intelligently tracks which memory pages are actively being used ("hot") and which are sitting idle ("cold"). Hot pages stay in DRAM for full-speed access. Cold pages get moved to NVMe storage, which is slower but orders of magnitude cheaper. The VM sees a single, larger memory space. It does not know the difference.

The economics are striking. A host with 1TB of DRAM can now present 2TB of total memory at the default 1:1 ratio, and up to 4TB at a 1:4 ratio for workloads with low active memory like VDI. Broadcom's own testing shows less than 5% performance loss when comparing 1TB DRAM-only to 1TB memory tiering configurations. For database workloads, they were able to double VM density per host with minimal performance impact. On a Dell PowerEdge R760, the per-server cost dropped from $55,878 to $33,792, a savings of nearly 40%. In a world where DRAM prices have surged 95% and memory now accounts for over half the cost of a new server, that is not a marginal improvement. It changes the procurement math entirely.

The feature also integrates natively with vMotion, DRS, HA, and vSAN. No special workflows. Most servers already certified for vSphere 8.x carry forward into VCF 9.0, so organizations can start using NVMe tiering on their existing fleet without waiting for new hardware that may take quarters to arrive.

The broader TCO impact compounds from there. Broadcom's VMware Telco Cloud Platform 9, built on VCF 9.0, claims 40% cumulative TCO savings over five years compared to siloed architectures. When server DRAM prices have surged nearly 95% and lead times stretch past a year, that kind of efficiency gain is not an optimization exercise. It is a strategic lifeline.

Nine of the top ten Fortune 500 companies have committed to VCF, with over 100 million cores licensed worldwide. The platform's appeal in this moment is clear: when you cannot buy your way out of a hardware shortage, you need software that makes every rack, every GPU, every gigabyte of memory count for more.

Broadcom is playing both sides of the board. On the silicon side, it is building the custom chips that offer a credible alternative to the NVIDIA GPU monoculture. On the platform side, it is giving enterprises the tools to stretch their existing infrastructure further. That dual position, chip architect and software platform, is unique in the industry. And in a hardware-constrained world, it may be the most valuable position of all.

The Haves and the Have-Nots

The most consequential near-term effect of the hardware shortage is a split in the AI world:

This is the dynamic that should concern strategists most. The AI revolution promises democratization of intelligence, but its infrastructure economics are pushing toward concentration. The companies that locked in forward GPU contracts in 2024 and 2025 have a multi-year structural advantage over everyone who did not. That advantage compounds. More compute means better models, which means more customers, which means more revenue to reinvest in compute.

What This Means for Enterprise Leaders

If you are an enterprise architect, CTO, or infrastructure leader, the hardware shortage demands a different kind of strategic thinking than the software era required. Software strategy was about talent and speed. AI strategy is about supply chain resilience.

Cyclical, Not Structural. But Respect the Cycle

History suggests the hardware shortage is not permanent. Every major technology bottleneck, from DRAM in the 1990s to hard drives after the Thailand floods to chip shortages during COVID, eventually resolved through a combination of investment, innovation, and demand moderation.

The AI hardware crunch will resolve too. Capital is flooding into new fabs. Alternative architectures are gaining traction. Model efficiency is improving faster than model size is growing. By 2028, the supply picture will look very different.

But "not permanent" does not mean "not consequential." The companies and strategies that are forged in this 2025 to 2028 window of constraint will define the competitive picture for a decade. Those who secure compute now, diversify their chip strategies, and invest in efficiency will have compounding advantages. Those who wait for the market to normalize will find that their competitors used the bottleneck as a forcing function to build better, leaner, more resilient infrastructure.

Software ate the world because it was free to distribute. AI is eating the world despite being expensive to run. That says something about how powerful the technology is. But it also means the winners of the AI era will not just be the ones who build the best models. They will be the ones who solve the hardware equation.