In-depth Industry Research Report · Computing power chips
Computing Power Chips 2026 Q2 In-Depth: Where Is Money Flowing? Where Is Price?
2026.08.07 · The full text is about 20 minutes · Data as of the close on 2026-08-06
AVGO · NVDA · AMD · MRVL · ARM · INTC
Key points of this article
1. The fastest-growing companies, AVGO (+143%) and NVDA (+92%), have underperformed the semiconductor sector by 37–41 percentage points since the beginning of the year.
2. INTC, whose growth rate is not disclosed, rose 170%—the market is buying narrative purity this year, not profit momentum.
3. In the past month, NVDA +11.2%, AVGO +13.4%, while the other four all declined: divergences have begun to narrow and are only just beginning.
Data as of the close on 2026-08-06. All financial figures are sourced from the original 8-K / 6-K filings submitted by six companies to the SEC, each with its source and background. This article is for research analysis and does not constitute investment advice.
First, let me tell you something that made me pause
Our peers' panel first concluded that computing chipsets are "at the bottom in NVDA growth rate."
This statement comes from a seemingly flawless table: AVGO +200%, MRVL +35%, AMD +26%, ARM +22%, NVDA +18%. Six companies, five numbers, NVDA ranked last.
Then I went to translate NVIDIA's original 8-K sentence:
Original sentences from the financial report
"First-quarter revenue was a record $75.2 billion, up 21% from the previous quarter and up 92% from a year ago."
NVDA's real year-over-year growth rate is +92%. The +18% in the table is the month-on-month figure for its data center computing business—the same sentence even says "Year-on-year +77%", but the machine mispicked the one that is.
What's worse, this is not an accident. NVDA's financial report shows four growth rates, and the machine captured the lower quarter-on-quarter figure four times:
Original sentences from the financial report
Total revenue: $75.2B
Year-on-year increase of +92%.
Month-on-month increase +21%.
Machine-grabbed +21%.
Data center calculations cost $60.4B
Year-on-year increase of +77%.
Month-on-month increase +18%.
Machine-grabbed +18%.
Data center network $14.8B
Year-on-year increase of +199%.
Month-on-month increase of +35%.
Machine-grabbed +35%.
Edge computing $6.4B
Year-on-year increase of +29%.
Month-on-month increase of +10%.
Machine-grabbed +10%.
On the same table, AVGO rose +200%, MRVL rose +35%, and ARM increased +22% year-on-year. As a result, a company with a 92% year-on-year growth was ranked at the bottom by its month-on-month figures alongside others' year-on-year figures.
I put this at the beginning, not to talk about a technical glitch. Rather, this happens to be the perfect metaphor for the mistake this article is discussing: the biggest pricing error in the computing chip sector right now is that the market uses different scales to measure different companies, and then lines up to pay according to the numbers you get from the volume.
(This extraction defect has been identified and fixed. I have checked every detail of the following figures back into the original financial report. ）
After getting the clock right, the two have completely different faces
The real growth rate of six companies
Company
AVGO
Latest disclosure as of 6/3 (FQ2).
Caliber AI Semiconductor revenue was $10.8B
Year-on-year growth +143% (Q3 guidance $16.0B, >+200%)
AMD
Latest disclosure as of 8/4 (Q2).
Caliber Data Center $6.718B
Year-over-year growth +107% (total revenue $11.536B, +50%)
NVDA
Latest disclosure: 5/20 (FQ1).
Caliber's total revenue was $75.2B
Year-on-year growth rate +92% (Data Center Network +199%)
MRVL
Latest disclosure as of 5/27 (FQ1).
Caliber total revenue $2.418B
Year-on-year growth +28% (Q2 guidance +35%, acceleration)
ARM
Latest disclosure on July 29
Total revenue of $1.289B
Year-on-year growth +22% (royalties +22%, licenses +23%)
INTC
Latest disclosure : 7/23
Caliber —
Year-on-year growth did not disclose revenue growth
Three caliber issues that must be clarified first
First, this is not the same quarter. The financial reports of the six companies spanned from June 3 to August 4, and NVDA's data is now 79 days outdated. Any interpretation of these six numbers as a "same-time comparison" is incorrect—it is a scrolling window.
Second, AVGO's +200% is guidance, not reality. What actually happened was +143% of Q2. But AVGO's acceleration trajectory has been the cleanest in these three quarters: +106% in Q1→ +143% in Q2→ >+200% in Q3, and the +143% in Q2 is what the company itself calls "above our forecast." This is a pattern of two consecutive upward revisions.
Third, MRVL is not slowing down. The panel shows it is "+35%, last period +42%," which looks like a 7 percentage point decline. But that +42% represents the growth rate for the entire FY2026 fiscal year, while +35% is the guidance for the next quarter. The real sequence is: FY2027Q1 Actual +28 % → FY2027Q2 Guidance +35%. The company's exact words over the phone were:
Original sentences from the financial report
"We expect year-over-year revenue growth to accelerate each quarter in fiscal 2027."
MRVL is accelerating, only up to 35%—still significantly below the top three. The direction is right, and the ranking hasn't changed.
Three Main Themes One: This is not a demand peak, but a budget redirection
The panel also concluded: "All four companies disclosing demand inflection points are marked as none, and no members in the group report a directional shift in demand."
This statement is also wrong, and even more thoroughly. The original text of the four companies is:
·AMD："Demand continued to strengthen across multiple end markets."(Demand in multiple end markets continues to strengthen)
·AMD："Customer forecasts exceeding our initial expectations."(Customer forecasts exceeded our initial expectations)
·NVDA："Computing demand is growing exponentially — the agentic AI inflection point has arrived."(The turning point for agent AI has arrived)
·MRVL："We expect year-over-year revenue growth to accelerate each quarter."
·AVGO："Our AI revenue growth is accelerating."
All four companies reported strong demand. The panel's statement that "no one reported the turning point" is actually the opposite of the fact.
But what really matters is where the money flows. Putting together four pieces of independent evidence:
Evidence figures
AVGO custom ASIC + AI network +143% (and guidance >200%)
NVDA's own network business up +199%.
NVDA's own general-purpose GPU calculation +77%.
AMD data center +107%.
Pay attention to the second and third lines—these are numbers from within the same company. NVIDIA's network business is growing 2.6 times faster than its GPU computing business. It's not that competitors are competing for its share; it's that its own revenue structure is drifting.
Three independent clues point to the same judgment:
Core judgment
Incremental computing power budgets are shifting from "buying general-purpose GPUs" to "buying custom ASICs + buying interconnects."
Why is that? The mechanism is actually quite simple. When a customer trains only a few models, buying the fastest general-purpose chip is the optimal solution—flexible, general-purpose, and mature ecosystem. But when a customer's workload is fixed and scaled to gigawatts (GW), two things happen simultaneously:
1. Custom chips start to become cost-effective. The one-time design cost of a dedicated ASIC for a fixed load is diluted by massive shipments, with the unit computing power cost significantly lower than that of general-purpose GPUs. This is AVGO's business.
2. Bottlenecks have shifted from 'how fast you calculate' to 'how fast you transmit information.' In a cluster of 10,000 cards, chips spend most of their time waiting for data. At this point, networks and optical interconnection become the key factors determining actual throughput. This is why NVDA Network +199% and why it signed multi-year optical agreements with Coherent, Corning, and Lumentum all at once.
So the question "Has AI demand peaked?" is a mistake. Demand hasn't peaked yet—four companies are saying it's accelerating. What has changed is where this money will be placed in the industry chain. General-purpose GPUs are still growing rapidly (+77% is not a weak figure), but it is no longer the fastest-growing segment.
Fourth, Main Theme Two: The capital returns of the two routes differ by an order of magnitude
The panel made a comparison: INTC's $5 billion capex versus AVGO's $231 million. This comparison is numerically invalid—INTC's $5 billion was a one-time expansion announcement (for Xeon 6 capacity on Intel's 3 process), while AVGO's $231 million was a quarterly actual capital expenditure. One is promise, the other is traffic—incomparable.
But if you look at a comparable caliber, the conclusion is even more glaring:
Core judgment
AVGO spent $ 231 million in FQ2 capital expenditures, generating $10.493 billion in operating cash flow and $10.262 billion in free cash flow.
Capital expenditures accounted for only 2.2% of operating cash flow.
This is the ultimate form of the fabless model: chip design, outsourcing manufacturing, and almost no capital pressure. INTC took the opposite path—building its own foundry (IDM), Panther Lake has entered mass production using ASML's High-NA EUV, and has also invested an additional 5 billion yuan in capacity expansion.
There is no right or wrong between these two routes, but they perform completely differently within the cycle:
·Fabless funds are mainly spent on R&D, allowing for rapid adjustments and disbursement as needed, with very strong cash flow resilience during downturns;
·IDM money is tied to production lines, which must be fully loaded once built to be cost-effective—so their tolerance for error in demand judgment is extremely low.
The practical meaning for investors is: In any scenario where AI demand slows down, AVGO's structure suffers much less cash flow loss than INTC. INTC is betting on a capacity cycle that will take years to validate, and it has not disclosed any revenue growth guidance.
This leads to the most important risk warning in this article, which I will elaborate on in Section Six.
5. The client list itself is a competitive ranking
The least numerical parts of a financial report often contain the most information. Let's list the partners named by the six companies together:
The competitor named in the company's financial report
AMD
Anthropic (up to 2GW MI450 Instinct + Helios rack), Meta (up to 6GW, first 1GW with custom MI450), Microsoft Azure (Helios + 6th generation EPYC), Cisco, Samsung (HBM4 supply)
NVDA
Google Cloud (Vera Rubin A5X instance), CoreWeave (5GW before 2030), Anthropic (investment + deep collaboration), Meta (multi-generational multi-generational), Coherent / Corning / Lumentum (optics), Marvell (NVLink Fusion + silicon photonics).
INTC
Fortinet (security processor), Foxconn (rack-level demo), ASML (High-NA EUV vendor), NVIDIA (appearing as a "vendor" in their own demo solution).
Three ways to read it:
First, AMD's order units are "gigawatts," not "pieces." Anthropic up to 2GW, Meta up to 6GW. This is the strongest fundamental shift in the entire group this quarter—a year ago, AMD was a "second supplier to watch" in accelerators, but now it has secured gigawatt-level commitments from two leading model companies. This is also the source of its data center revenue of +107% year-on-year. (Note: These are the caps of the framework agreement, not confirmed order books.) ）
Second, NVDA's partnership with Marvell reveals its true defense direction. NVLink Fusion opens NVIDIA's interconnect protocol to third-party chips—a company building a wall on its fastest-growing segment (network +199%), rather than stubbornly sticking to the slower-growing GPU. Management saw the budget reversal earlier than the market.
Third, NVIDIA appears on the INTC list as a supplier. In Intel's own 8-K, the "Disconnected Agent Cloud" configuration is Intel Xeon + SambaNova RDU + NVIDIA Blackwell. In other words, in Intel's own AI story, the chip responsible for computing power is not Intel's.
6. The real misalignment: profits are on one side, stock prices are there
Up to this point, all are financial reports. Now let's see what the market is paying for.
Performance of six companies from 2026 to date (as of the close of 8/6):
Company
INTC
Year-to-date +170.5%.
In the past month , −9.6%.
Closing close since 2025/11: −29.2%.
Year-on-year growth rate not disclosed
ARM
Year-to-date +162.3%.
In the past month , −4.6%.
Highest closing since November 2025: −34.8%.
Year-on-year growth rate +22%.
MRVL
Year-to-date +147.9%.
In the past month , −8.7%.
Closing close since 2025/11 −33.5%.
Year-on-year growth rate +28%.
AMD
Year-to-date +128.5%.
Nearly 1 month : −5.2%.
Closing close since November 2025: −15.8%.
Year-on-year growth rate +107% (data centers).
AVGO
Year-to-date +21.5%.
+13.4% in the past month
Closing down −12.7% since November 2025
Year-on-year growth rate +143%.
NVDA
Year-to-date +17.4%.
+11.2% in the past month
Highest closing since November 2025: −7.1%.
Year-on-year growth rate +92%.
Semiconductor ETF (SMH)
Year-to-date +58.7%.
Nearly 1 month : −1.7%.
Closing close since November 2025: −14.6%.
S&P 500 (SPY)
Year-to-date +12.7%.
+2.8% in the past month
Highest closing since November 2025: −0.4%.
If you look at this table side by side with the previous one, you'll notice a fact that's hard to ignore:
Core judgment
The two fastest-growing companies (AVGO +143%, NVDA +92%) have lagging the semiconductor sector by 37 to 41 percentage points since the beginning of the year.
The company with the lowest growth rate (INTC), which does not disclose revenue guidance at all, rose the most—+170%.
This is almost a perfect negative correlation. This year, the market is paying for computing chips in the opposite direction of the company's reported growth direction.
Why is that?
My explanation is: the market is not buying profit momentum this year, but the narrative of "AI purity."
·INTC is selling a "revival of advanced American processes + foundry turnaround"—a grand story that cannot be falsified by quarterly numbers. Precisely because it does not disclose revenue guidance, the story is not interrupted by numbers.
·ARM sells "all AI chips have to pay me royalties"—a structured story that sounds unavoidable. However, its actual revenue growth rate was +22%, the lowest effective figure in the group.
·MRVL sells the "second supplier of custom ASICs"—real but with a +28% growth rate.
·What about AVGO / NVDA? They are too big, too well understood, and the story is finished. No narrative premium to pay.
Narrative pricing has a characteristic: it remains valid until it is cashed out and loses its effect the moment it is tested. So we see the third column—
The retraction column is evidence that the narrative is being withdrawn
·ARM −34.8%、MRVL −33.5%、INTC −29.2%
·AMD −15.8%
·AVGO −12.7%、NVDA −7.1%
The three biggest gains also saw the deepest drawdowns; The one that fell the least was NVDA, which underperformed the most throughout the year. In the past month: NVDA +11.2%, AVGO +13.4%, while the other four all declined, with SMH itself down −1.7%.
This isn't random fluctuation; it's a directional rotation: money is moving from narrative stocks back to cash flow stocks. And this rotation has only been underway for a month, and the 37–41 percentage point gap between NVDA and AVGO and the sector has only been recovered by a small portion.
This is the core judgment of this article: within the computing chip sector, there is an extreme divergence between profit momentum and stock price momentum, and the divergence has begun to narrow. This convergence process itself is the best odds in the current sector.