Tonight's Broadcom earnings report isn't really about how much money it made, but whether Wall Street will officially recognize that the AI chip battlefield has moved from NVIDIA's dominance to a new phase of "GPU plus custom chips plus storage plus networking." More bluntly, NVIDIA is still the king of AI arms dealers, but Broadcom isn't grabbing the crown—it's grabbing the printing press next to the throne.
We don't chase the loudest hype, only the most valuable logic. Today's episode focuses on one question: Can Broadcom's earnings tonight ignite storage? Is AI custom chip really going to steal NVIDIA's lunch?
I don't think Broadcom will immediately replace NVIDIA. NVIDIA's GPU ecosystem, CUDA software moat, and whole-system capabilities remain core for AI training and high-end inference. But I believe the AI investment logic for 2026 has shifted. The market used to buy "who can sell the strongest AI chip"; now it's starting to buy "who can help big tech lower AI costs, scale up, and keep profits."
This change is critical. AI is no longer a few scientists in a lab tuning models with headphones; it's millions of users worldwide asking Claude, Gemini, and ChatGPT every day. Enterprises run AI customer service, AI coding, AI search, and AI office daily.
So, the real conflict is here.
NVIDIA sells top-tier general-purpose GPUs—like a Swiss Army knife that can do everything and do it well. Broadcom's custom chips are like bespoke kitchen assembly lines for big tech: they only make fixed dishes, but they're cheap, efficient, and have clear cost advantages at scale.
That's why tonight's Broadcom earnings are so important.
It's not just a chip company's quarterly report; it's Wall Street checking a bigger question: Will AI giants like Google, Anthropic, Meta, Amazon, and OpenAI continue to buy NVIDIA GPUs indefinitely, or will they shift more spending to self-developed chips, custom ASICs, TPUs, AI networking switches, HBM high-bandwidth memory, and enterprise SSDs?
In the past, AI discussions were inseparable from NVIDIA. Now, money is spreading to the periphery. Broadcom rises, Marvell rises, Micron rises, SanDisk rises, SK hynix rises, Samsung gets revalued. Previously, AI markets were like a highway with all cars crowding NVIDIA's toll booth. Now it's like a large airport requiring not just runways but control towers, fuel, baggage systems, security, and power. NVIDIA is still the jet engine, but Broadcom could be the control tower and dedicated runway, while storage companies are fuel and warehouses.
Let's look at Broadcom's earnings expectations tonight.
Market focus isn't just total revenue, but whether AI semiconductor revenue can continue to explode. Last quarter, Broadcom's total revenue was $19.311 billion, up 29% YoY, with AI semiconductor revenue reaching $8.4 billion, up 106% YoY. What does that mean? Broadcom was a large factory selling various high-end components; now one AI production line has become a cash cow—and that cow is drinking protein powder.
The company's guidance for this quarter is around $22 billion, nearly 47% YoY growth. That's not typical tech growth—it's the kind of story Wall Street loves: old businesses remain, software provides profit base, and AI business offers valuation imagination.
But precisely because the stock has surged so much, tonight's earnings bar is set very high.
If Broadcom is only "pretty good," the market may not be satisfied. Current valuations aren't rewarding "pretty good"—they're pricing in "you must be very strong." With Broadcom's stock near historical highs, the market isn't asking whether it's growing, but whether growth can accelerate further, AI orders can be seen further out, gross margins can hold, and customer concentration risk won't increase.
So why is Broadcom being revalued by the market?
Core: three words—custom chips, AI networking, long-term customer lock-in.
Let's start with custom chips. Many think custom chips are extremely sophisticated. In plain terms, big tech doesn't want to buy others' general-purpose chips forever, so they partner with companies like Broadcom that have chip design and supply chain experience to create dedicated AI chips better suited to their models and workloads.
Google's TPU is a typical example. Google has been working on TPUs for years, with Broadcom participating behind the scenes. Previously, Google was seen as a search, advertising, and cloud company. In the AI era, Google has an easily underestimated identity: it's one of the few super-players that has long self-developed AI chips and integrated them into its cloud services.
AI giants fear not chip shortages but long-term compute cost loss of control. Think about it: more Claude users, more Gemini users—every conversation, code generation, long-text analysis essentially burns electricity, servers, and chip depreciation. If fully dependent on external expensive GPUs, faster user growth means scarier bills. On the surface, user numbers explode, but the CFO might be silently crying in the office.
So the partnership between Anthropic, Google, and Broadcom triggers a big market reaction. Anthropic has announced expanded use of Google Cloud and TPUs, obtaining multi-gigawatt-level TPU compute power starting in 2027. "Multi-gigawatt" isn't an ordinary news term—it represents a hyperscale AI infrastructure commitment.
What's a gigawatt? Not one or two servers, not a small computer room—it's near large power-project-level compute deployment. AI has shifted from "grabbing chips" to "grabbing power, land, cooling, networking, storage, and long-term supply chains." That's why Broadcom, storage, and networking chips suddenly become so attractive.
Because AI is no longer a stage where you can tell a story with a few GPUs. Now entire data centers must be rebuilt.
This is also the new knowledge I want to share today that few other bloggers explore: the real underlying conflict of the AI chip war isn't which is stronger—GPU or ASIC—but the war between "generality" and "unit economics."
NVIDIA's GPU is strong in generality. When models are still changing rapidly, big tech buys GPUs for flexibility, mature ecosystem, many developers, suitable for training and complex inference. But when certain model structures, inference scenarios, and internal workloads stabilize, the value of dedicated chips emerges.
For example, an AI company handles massive similar tasks daily—customer service, search ranking, recommendations, speech recognition, code completion, ad generation. It doesn't need the most expensive universal tool for every task; it needs to push high-frequency tasks to the lowest cost.
The AI inference era is like that. Training stage is about who can build the model; inference stage is about who can run the model cheaply. NVIDIA remains important, but ASIC and TPU importance will grow. Broadcom's opportunity lies here.
The second line is AI networking. Many think an AI data center is just a bunch of GPUs together. Actually, the real challenge is making these chips talk to each other like one brain. If chip-to-chip communication is too slow, data gets stuck on the road—even the strongest chip becomes a luxury car stuck on LA freeways.
Broadcom is very strong in Ethernet switching chips, like the Tomahawk series, a key device in AI data center networks. Larger AI models, more parameters, higher inference concurrency mean more terrifying data flow between chips, servers, and racks. At this point, networking isn't a supporting role but a bottleneck.
That's why Broadcom's AI story isn't about a single chip—it's a combination of "custom computing plus network connectivity." It can participate in customers' custom AI chips and also sell data center networking chips. The more AI clusters customers build, the more slices Broadcom can eat.
Note: this is also why Broadcom and Marvell are simultaneously sought after by the market. Marvell was recently praised by NVIDIA CEO Jensen Huang, directly exciting the market. Jensen's one sentence is like a Michelin review in the AI world—before the chef even serves the dish, the stock price is already on the table.
But we must stay calm. Behind this isn't simple hype-grabbing; AI infrastructure is spreading from "compute chips" to "connectivity chips." As compute clusters grow larger, networking, optical interconnects, switching chips, custom chips, memory, and storage all become key links. Whoever can reduce latency, lower power, and increase throughput gets big tech budgets.
The third line is storage.
I want to emphasize today: many look at Broadcom only to see if it will steal NVIDIA's lunch. But I think more easily underestimated is Broadcom's earnings verification role for the storage sector.
Why?
Whether it's NVIDIA GPUs, Google TPUs, or Broadcom-involved custom ASICs, they all ultimately need one thing: data. AI doesn't think in a vacuum; it needs to read data, store data, adjust parameters, cache context, handle vector databases, support long texts, support multimodality. The stronger the compute, the greater the data movement pressure. The smarter the model, the more terrifying the memory demand.
In the past, the market understood AI chips as "computing power." But in 2026, it's increasingly clear that AI's bottleneck is shifting from pure compute to memory bandwidth and storage capacity.
Why is HBM high-bandwidth memory expensive? Because AI chips don't just need to compute; they need to feed on data. The chip is like a super chef; HBM is the conveyor belt constantly delivering ingredients. No matter how great the chef, if the belt jams, you're just staring at the pot in the kitchen.
That's why storage companies like Micron, SK hynix, and Samsung are being revalued. Especially after HBM demand explosion, the storage industry has seen an interesting shift: it used to be a typical cyclical stock—prices rise when everyone profits, then expansion leads to price collapse and industry enters hell mode. But this time is different: AI demand is longer-term, customers are more concentrated, suppliers expand more cautiously, advanced packaging and HBM capacity ramps are harder, making supply-demand tightness potentially more persistent than past cycles.
Of course, I'm not saying storage stocks only go up. Storage stocks rise fiercely and fall unkindly—they're not milk tea, but baijiu: fragrant on entry, strong aftertaste. But within the AI infrastructure cycle, storage's status is indeed higher than before.
Especially HBM and enterprise SSDs.
HBM serves the data bandwidth inside AI accelerators; enterprise SSDs serve data calls in data center storage and model inference. In the future, as AI agents multiply, enterprise documents, codebases, customer data, images, videos, voice data must be vectorized, indexed, and called. You think AI just answers a sentence, but behind it, it might rummage through a huge database for information.
So I'm bullish on storage, not because it has risen a lot, but because the AI inference era's demand for "memory" and "data channels" may be longer than the market imagines.
That's today's core investment framework. NVIDIA represents the strongest compute; Broadcom represents customization and networking; Marvell represents connectivity and some custom chips; Micron, SK hynix, Samsung, SanDisk, Western Digital represent storage and data carriers; TSMC represents manufacturing and advanced packaging. AI isn't a single stock trade; it's a revaluation of an entire industry chain.
Having said that, we must also discuss risks.
Broadcom isn't an obscure small-cap that no one has discovered. It's already a market focus; the stock price has priced in many optimistic expectations. If tonight's earnings don't deliver "beat and raise again," short-term volatility will be high. Custom chips sound beautiful, but they naturally depend on large customers. The capex pace of customers like Google, Anthropic, Meta directly impacts Broadcom's AI revenue. If a major customer delays a project, changes architecture, or pushes for lower prices, Broadcom will be affected.
ASICs are efficient but less flexible than GPUs. If AI model architectures change rapidly, or new models require different computation methods, custom chips may face iteration pressure. Big tech customizes because they have scale, engineering capability, and long-term needs. Ordinary companies may not be suitable.
At this point, some friends might ask: what does this mean for ordinary investors?
I'll give you a practical framework: don't just ask "should I buy Broadcom?" Ask yourself which part of the AI industry chain you want to buy.
If you want the highest-certainty AI compute leader, NVIDIA remains core, but valuations and expectations are high.
If you want AI custom chips and networking expansion, Broadcom is key to watch; Marvell is an important comparator.
If you want data bottlenecks in the AI inference era, Micron, SK hynix, Samsung, SanDisk, Western Digital and similar storage chains are worth tracking.
If you want to buy the picks-and-shovels of picks-and-shovels, TSMC, advanced packaging, semiconductor equipment, power and cooling infrastructure, all have long-term logic.
But my personal preference: don't go all in when market sentiment is hottest around earnings. Instead, treat tonight's earnings as an industry chain checkup.
If Broadcom earnings and guidance are strong, AI custom chip and networking demand continue accelerating. Then on pullbacks, focus on the AI infrastructure chain, not just the hottest K-line.
If Broadcom earnings are decent but the stock opens high and goes low, it doesn't necessarily mean the thesis is broken—short-term expectations may be too full. Then listen to the management call for keywords: order visibility, AI revenue growth, customer expansion, networking chip demand, gross margin changes.
If Broadcom earnings clearly miss expectations, be careful with short-term valuation pressure across the AI infrastructure chain, especially storage and networking chips that have risen sharply—they could pull back together.
So tonight's Broadcom earnings aren't just about whether Broadcom rises; it's about Wall Street confirming whether AI markets are entering a second phase.
Phase 1: Buy NVIDIA, buy the strongest compute.
Phase 2: Buy AI infrastructure, buy custom chips, networking, storage, power, cooling, packaging—all the key links that enable AI to move from lab to commercial-scale operation.
If this thesis holds, the real things worth studying next aren't who shouts loudest, but who is landing long-term orders from big tech, who is solving AI cost issues, and who is grabbing the real bottlenecks in data centers.
Tonight's Broadcom earnings are an important battle to test this logic.
