The AI Infrastructure Trade Isn't About Earnings — It's About Where You Sit When Clusters Scale

On August 4, AMD reported Q2 revenue of $11.5 billion — up 50% year-over-year — with Data Center sales more than doubling to $6.7 billion. Non-GAAP EPS more than tripled to $1.66. By every metric on a standard "strong earnings, healthy balance sheet" screen, the company checks both boxes. Its net cash position sits at nearly $10 billion, debt-to-equity is 4.8%, and free cash flow growth has accelerated to 108% year-over-year.

The stock fell more than 10% after the close anyway.

That reaction is the signal I pay attention to. The narrative around AMD as Nvidia's dark horse has already propelled its market cap to $790 billion and its stock up 125% year-to-date. When a company beats a widely beaten-up earnings estimate and the market sells it off, it means the thesis has been fully digested. The remaining question is no longer "is this company growing" — it's "is the return profile still compelling relative to what's available elsewhere in the AI trade?"

That question doesn't get answered by a generic earnings-growth-and-balance-sheet screen. It gets answered by looking at where each company sits in the architecture of the AI buildout. The five largest hyperscalers — Microsoft, Amazon, Alphabet, Meta, and Oracle — are on track to spend as much as $750 billion on capital expenditure this year. That figure is projected to climb above $1 trillion in 2027. The question is which part of that spending machine carries the best combination of growth, margin quality, and financial leverage.

I'm looking at three companies at the core of this transition: Arista Networks, Broadcom, and AMD. All three are growing rapidly. All three have balance sheets that pass a basic stress test. But only one of them sits in the part of the stack where the margins, the cash flow quality, and the leverage risk create the most defensible setup.

Arista: The Near-Debt-Free Play on the One Bottleneck Everyone Acknowledges

Arista Networks delivered its first-ever $3 billion quarter in Q2 2026, with revenue of $3.04 billion, up 37.7% year-over-year and 12.1% sequentially. The company beat consensus on both revenue ($2.83 billion expected) and non-GAAP EPS ($1.02 actual versus $0.89 expected). Then it raised full-year guidance.

Q3 guidance sits at approximately $3.3 billion, with non-GAAP operating margins of 48% to 49%. That's acceleration in a business that's already growing in the high-30s.

Here's where the balance sheet actually matters. Arista carries $8.9 billion in total debt against $2.3 billion in cash and $11.1 billion in marketable securities — effectively net cash. Its current ratio is 296%, its debt-to-equity is zero percent (the system registers it as negligible against total equity of $14.8 billion), and its free cash flow margin TTM is 54.4%. Return on invested capital sits at 28%. Those aren't software-company metrics — they're a hardware business running at software-quality returns, because Arista doesn't carry the fab risk, wafer cost swings, or advanced packaging capital expenditure that its semiconductor peers do.

The product positioning makes the financials make sense. Arista's new 7060XE7 Series delivers 1.6 terabits of switching capacity with up to 100 Tbps of system bandwidth. Its Linear Pluggable Optics technology reduces interconnect power consumption by approximately 60% compared to traditional pluggable optics. CEO Jayshree Ullal calls this the "Arista 2.0 platform strategy" — positioning networking as the central nervous system connecting client to campus to AI data center.

Put plainly: as AI clusters grow from thousands of GPUs to tens of thousands, the bottleneck shifts from compute to interconnect. Every Nvidia H200, every AMD MI450, every Google TPU needs to talk to every other chip in the rack, and across racks, at speeds where milliseconds matter. Arista sells the pipes. It doesn't need to win the chip war. It just needs to be the vendor that hyperscalers trust when failure means millions of dollars in idle compute.

The company is that vendor. A 2026 Henry Fund analysis describes Arista as "the dominant Ethernet switching vendor for hyperscale AI infrastructure, capturing disproportionate share" of a $20.8 billion cloud AI networking market. Gartner named it a Leader in its 2026 Magic Quadrant for enterprise networking.

Broadcom: The Custom Silicon Machine With a $91 Billion Question Mark

Broadcom operates on an entirely different scale. The company posted 32.3% revenue growth year-over-year, with AI revenue more than doubling in recent quarters. CEO Hock Tan has drawn a line of sight to "significantly in excess of $100 billion" in AI chip revenue by 2027.

The customer list reads like a roll call of the companies writing those $750 billion capex checks. Google's seventh-generation TPUs, Anthropic scaling from 1 gigawatt of compute to over 3 gigawatts in 2027, Meta's MTIA custom accelerator program (which Tan insists is "alive and well"), and OpenAI deploying its first custom chip at over 1 gigawatt in volume in 2027. Tan has secured memory capacity through 2028, which matters enormously in a market where high-bandwidth memory shortages have already delayed GPU shipments.

Broadcom's networking business is also accelerating. Its Tomahawk 6 switch chip delivers 100 terabits per second, with Tomahawk 7 — doubling performance — expected in 2027. AI networking revenue grew 60% year-over-year and is projected to grow to 40% of total AI revenue. VMware, acquired for $69 billion, added $6.8 billion in Q1 revenue with 19% annual recurring revenue growth. Tan's thesis that VMware is the "permanent abstraction layer" between AI software and physical silicon gives Broadcom a software-recurring anchor that Arista doesn't have.

The financial profile is impressive by most standards. Operating margin of 43.4%, free cash flow margin of 43.4%, return on invested capital of 21.8%. Free cash flow TTM totals $32.8 billion.

But total debt is $91.5 billion. Net debt — debt minus cash — is $45.3 billion. Debt-to-equity sits at 74%. The company pays out $2.53 per share annually in dividends and returns capital through buybacks, all while financing an aggressive acquisition strategy. Put it against the backdrop of Hock Tan's $100 billion AI revenue target, and the question isn't whether Broadcom is growing. The question is whether that leverage becomes a constraint when memory prices spike, advanced packaging capacity tightens at TSMC, or a single hyperscaler customer scales back its custom silicon program.

I believe Broadcom's thesis is structurally sound. The custom silicon transition is real, and Broadcom has 20 years of experience in high-volume silicon production with few competitors who can match its yield and time-to-market record. But the return on that thesis has been front-loaded. The stock is up 40% over the trailing year, trading at a market cap of $2.04 trillion. At those levels, the margin for error narrows.

AMD: The Dark Horse That Just Became the Consensus

AMD's Q2 results were genuinely strong. Revenue of $11.5 billion, up 50% year-over-year. Data Center revenue of $6.7 billion, up 107% year-over-year, now representing 58% of total company revenue. The company launched its Helios rack-scale AI server, with deployments announced at Anthropic, Meta, Microsoft, and OpenAI. CEO Lisa Su told analysts that AI is "driving a significant expansion in demand for compute across all of our markets" and that AMD's "leadership portfolio and growing customer visibility position us exceptionally well."

Q3 guidance of approximately $13 billion implies roughly 41% year-over-year growth. The company's balance sheet is pristine — $5.1 billion in cash, net cash position of $9.9 billion, 108% free cash flow growth. FCF margin sits at 22.9%.

But the 10% post-earnings sell-off tells you everything you need to know about the current setup. AMD has nearly tripled over the past year. At $790 billion in market cap, the stock trades at 19.1 times trailing sales — a multiple that assumes the MI400 series ships on time, Helios captures meaningful share against Nvidia's rack solutions, and the Anthropic 2-gigawatt deployment actually materializes. The market has already priced in the dark horse narrative. When the narrative becomes the consensus, the return curve flattens.

The architecture story remains compelling. AMD's MI450 Series GPUs and 6th Gen EPYC processors represent a genuine generational gap on paper, and the move to rack-scale solutions with Helios is the right architectural direction. But execution risk in semiconductor hardware is real, and the company's operating margin of 11.7% — while improving — is still a fraction of what Arista and Broadcom deliver. That gap has to close for the market to justify the current multiple on sustained basis.

Where the Capital Goes

The competitor headline asks a screening question. I ask an allocation question. All three companies are growing. All three have defensible balance sheets. But the combination of growth trajectory, margin quality, financial leverage, and valuation compression points in one direction.

Metric	Arista	Broadcom	AMD
Revenue Growth (YoY)	32.6%	32.3%	39.5%
Operating Margin	42.8%	43.4%	11.7%
Free Cash Flow Margin	54.4%	43.4%	22.9%
Net Debt Position	~$0 (net cash)	$45.3B	$9.9B (net cash)
ROIC	28.0%	21.8%	8.1%
Market Cap	$238B	$2.04T	$790B
P/S (TTM)	22.6x	27.0x	19.1x

Arista trades at the smallest market cap of the three — $238 billion — despite posting the highest ROIC, the strongest free cash flow margin, and essentially no debt. AMD trades at nearly four times Arista's market cap with a fifth of the ROIC and operating margins that are one-quarter of Arista's. Broadcom trades at nearly nine times Arista's market cap while carrying $45 billion in net debt.

The market has priced the GPU story. It has priced the custom silicon story. It hasn't fully priced the networking bottleneck story.

I believe Arista deserves the largest allocation of the three at current levels. The company sits at the intersection of three durable trends: AI cluster scale expanding from thousands to tens of thousands of GPUs (which exponentially increases networking demand), power constraints making efficient interconnects more valuable, and the shift from proprietary InfiniBand fabrics to open Ethernet architectures where Arista dominates. The balance sheet gives it optionality — it can invest through a downturn, acquire complementary technology, or return capital without stress.

Broadcom remains a compelling long-term hold for investors who believe the custom silicon transition accelerates faster than consensus expects. But the debt load and the $2 trillion valuation compress the near-term return profile. In my opinion, Broadcom is better suited to a core position bought on weakness than chased at current levels.

AMD is a stock I believe in but wouldn't add to at the current price. Much of its return curve is back-half weighted — the Helios deployments, the MI400 ramp, the Anthropic gigawatt-scale partnership — all of that needs to materialize before the $790 billion market cap looks justified. A pullback creates a setup. The current price reflects too much of that future in advance.

The debate isn't about whether these three companies matter to the AI buildout. They do. The debate is about which one gives you the best combination of growth, margin quality, and financial optionality at the price you're paying. For me, right now, that's Arista.

What Would Change My View

I'd reduce my Arista allocation if the company misses Q3 guidance of $3.3 billion or if operating margins compress below 45%, which would signal pricing pressure from Nvidia's competing InfiniBand networking or Broadcom's Tomahawk 7 switch chip arriving ahead of schedule. I'd add to AMD if the stock pulls back to the $350-$400 range — a 25-30% decline — while its MI400 and Helios execution remains on track. I'd increase my Broadcom conviction if Hock Tan's $100 billion AI revenue target is met by early 2027 and the company uses the resulting cash flow to materially reduce its $91 billion debt load. Until then, the leverage stays on my watch list.