SharkWater Trading  •  AI Infrastructure • Semis • Counterparty Risk

ORCL, MRVL, CBRS: Three Floors Of The Same Building, And Two Of Them Have The Same Landlord

August 17, 2026

Bottom Line Up Front

Oracle rents the building. Marvell sells the wiring. Cerebras builds a specialized engine that goes inside. All three are levered to the AI capacity buildout, but they are not three separate bets.

Roughly half of Oracle's $638 billion backlog traces to OpenAI. Roughly eighty percent of Cerebras' $25.4 billion backlog traces to OpenAI. Owning both is not diversification. It is the same counterparty bet placed twice, at two very different valuations. Marvell is the odd one out, and that is the most interesting thing about it.

The Stack At A Glance

	ORCL	MRVL	CBRS
Role	Capacity landlord	Interconnect and custom silicon	Wafer-scale inference
Price (Aug 17)	~$147.57	~$234.20	~$251.98
Market cap	~$420B	~$199B	~$60B
Revenue base	$67.4B FY26	~$10.8B run rate	$880 to $890M FY26 guide
Roughly, price to sales	~6x	~18x	~68x
Backlog (RPO)	$638B	Not reported this way	$25.4B
Backlog vs revenue	~9.5x	n/a	~29x
From 52-week high	Down ~57%	Down ~29% from $329	Down ~35% from $386
Next catalyst	FQ1 2027, ~Sept 8 to 14	FQ2 2027, Aug 27	SUPERNOVA event

Read the price to sales row from left to right and you are reading an escalator of expectations. Same theme, three wildly different prices for exposure to it.

Oracle: A Great Business Attached To A Terrifying Cash Flow Statement

Fiscal Q4 2026 was, on the surface, one of the better quarters any megacap has printed. Total revenue $19.2 billion, up 21 percent. Cloud revenue $9.9 billion, up 47 percent. Infrastructure alone $5.8 billion, up 93 percent. Full year revenue $67.4 billion, up 17.4 percent, with net income up 36.5 percent.

And RPO, the contracted backlog, hit $638 billion, up 363 percent year over year and up roughly $85 billion in that quarter alone.

The stock is down about 57 percent from its high.

Why The Market Stopped Caring About The Backlog

Because backlog is a promise, and promises have to be financed before they become cash.

Metric	FY2026
Capital expenditure	~$55.6B (above the $50B target)
Free cash flow	-$23.7B
FY2027 capex guidance	Up to ~$95B, roughly $70B net Oracle spend
Corporate debt	Above $110B

For context, Oracle averaged about $1.7 billion in annual capex between fiscal 2018 and 2021. It is now spending that much roughly every eight days.

The catalyst chain that broke the stock was specific. December's Q2 print disclosed $12 billion of capex in a single quarter and a $15 billion raise to the full year forecast, and shares fell 15 percent on the day. Blue Owl Capital then walked away from a $10 billion data center financing deal, reportedly over Oracle's spending pace and debt load. S&P downgraded. Securities class actions followed. And reports surfaced that OpenAI, the counterparty behind an estimated $300 billion of that backlog, had missed internal revenue targets.

Management has an answer, and it is a decent one: much of the recent RPO growth came from contracts where the customer either prepaid for the GPUs or supplied the GPUs directly, which reduces what Oracle itself has to finance. That is a real mitigant and it deserves weight.

The right analogy is a contractor who has just signed the largest order book in the history of the trade, from a client whose own financing is not fully secured, and who must now borrow heavily to buy the materials. The order book is real. The question is who runs out of money first.

Bull: 44 analysts carry an average target near $246, roughly 68 percent above the current price. The underlying database and applications business still throws off cash. If even half of that backlog converts on schedule, the revenue math is transformative, and the stock is priced at roughly six times sales for it.

Bear: Every dollar of that backlog assumes the customer can pay and Oracle can build. Gross margin has already compressed from 81 percent to around 69 percent as revenue mix shifts toward infrastructure. Negative free cash flow of $23.7 billion in a single fiscal year, funded by debt, in a business where the largest customer is a company that is itself not yet reliably profitable, is the definition of a leveraged bet on somebody else's business plan.

Marvell: The Only One Here Selling To Everybody

Marvell is up roughly 161 percent in 2026, joined the S&P 500 in June, took a $2 billion investment from NVIDIA, and got publicly called a future trillion dollar company by Jensen Huang. Q1 FY2027 revenue was $2.42 billion, up 27.6 percent, with Q2 guided to about $2.7 billion. Roughly 75 percent of revenue is tied to data center and AI demand.

It also trades around 81 times trailing earnings and about 18 times sales, and it peaked at $329 in June before falling 35 percent and clawing back to the mid $230s. This is not a cheap stock and it is not a calm one.

The Narrative Gap Worth Watching

Most of the retail conversation treats Marvell as a custom AI silicon story, a Broadcom competitor. On that axis the position is honestly mediocre. Marvell holds an estimated 20 to 25 percent design share against Broadcom's roughly 70 percent, and Broadcom serves three of the five major hyperscaler programs including Google's TPU, Meta's Iris, and OpenAI's Jalapeño. Marvell has Amazon Trainium and Microsoft Maia 200.

But management has guided optical interconnect revenue to grow above 70 percent in fiscal 2027, versus 20 percent plus for custom silicon. In optical DSPs, Marvell holds an estimated 60 to 65 percent share, built on a position that predates the AI boom entirely. The Celestial AI acquisition added plasmonics-based silicon photonics and XConn added PCIe and CXL switching.

The company's own guidance says the growth engine is the interconnect business, not the ASIC business the market argues about. If you are underwriting this name, that is where the work should go.

Bull: Every AI cluster needs optics regardless of whose accelerator wins. Revenue is recognized on shipment, not promised over eight years. Marvell has no OpenAI counterparty concentration, and its custom silicon customers are Amazon and Microsoft, two of the best-capitalized buyers on earth. Analyst consensus is Strong Buy with an average target near $257.

Bear: 81 times trailing earnings after a 161 percent run leaves no margin for a soft quarter, and one downgrade to Hold already flagged exactly that. Operating margin sat at 14 percent in Q1, in line with a year prior, meaning the revenue growth has not yet produced the operating leverage the story promises. Marvell is materially smaller than Broadcom and NVIDIA and has to outspend its weight class to stay relevant. Earnings land August 27, one day after NVIDIA, in the middle of the most concentrated AI sentiment window of the year.

Cerebras: Read The Warrant Footnote

Cerebras went public in May 2026, raising $6.4 billion in the largest semiconductor IPO on record, oversubscribed roughly twenty times. It ran to $386.34 intraday, collapsed to $160.81 by late June, and now sits around $252. Reported beta is above 5. Treat every number in this section as attached to a stock that moves 15 percent on a Monday.

The Q2 print on August 12 was, operationally, excellent. Core revenue $210 million, up 103 percent, against consensus of $193.6 million. Cloud revenue nearly quadrupled. Full year core revenue guidance was raised to $880 to $890 million from $855 to $865 million, and full year gross margin guidance was raised to 41 to 43 percent from 38 to 41 percent.

The stock fell 13 to 20 percent.

Here Is Why, And Almost Nobody Explained It Properly

In December 2025 Cerebras signed a master relationship agreement with OpenAI covering 750 megawatts of inference capacity through 2028, valued above $20 billion, with an option for an additional 1.25 gigawatts by 2030. OpenAI also provided a roughly $1 billion working capital loan in January 2026 and received warrants for approximately 10 percent of the company.

Under ASC 606, warrants issued to a customer as part of a commercial arrangement are treated as consideration paid to that customer. The fair value sits on the balance sheet as a customer warrant asset and gets amortized against revenue as the customer buys compute.

As of June 30, Cerebras carried roughly $1.128 billion of customer warrant assets, split $167.8 million current and $960.6 million non-current, and the 10-Q states this amortizes through October 2031.

Translation: there is a permanent, structural wedge between the "core revenue" management guides to and the GAAP revenue that hits the income statement, and it runs for another five years. If your data feed compares management's core guidance to a GAAP consensus, you will get a phantom miss every single quarter. That is roughly what happened on August 12, and it produced a $13 million GAAP shortfall that knocked a fifth off the market cap on a quarter where the business beat and raised.

Then on August 17 the stock ran 15 percent on renewed attention to the OpenAI relationship. Same company, same footnote, opposite direction, five days apart.

Bull: Independently benchmarked as the fastest inference hardware available, with an anchor contract above $20 billion, AWS distribution through Bedrock, and over $10 billion of capital to deploy. RPO of $25.4 billion against an $885 million revenue year, with roughly 22 percent, about $5.59 billion, expected to convert within 24 months. Margin guidance was raised, not cut.

Bear: Roughly eighty percent of that backlog is one customer, and the rest is concentrated in AWS and G42. The company is not profitable. It trades near 68 times current year revenue. Any slippage in OpenAI's deployment schedule does not dent the story, it removes it. And OpenAI holds warrants on ten percent of the company, which is an alignment mechanism and a dilution overhang in the same instrument.

The SharkWater Take

The concentration point is the one I would tape to the monitor. If OpenAI is more than half of Oracle's backlog and roughly four fifths of Cerebras' backlog, then a position in both is one position in OpenAI's ability to fund what it has signed, expressed through two balance sheets you do not control. Anyone holding both and feeling diversified should size accordingly.

On the individual names:

Oracle is the only one of the three where the pessimism is already in the price. Down 57 percent, roughly six times sales, with a real underlying enterprise business underneath the AI wager. It is also the one with $110 billion of debt and negative $23.7 billion of free cash flow, so "cheap" is doing a lot of work. The September print is the tell, and the numbers that matter are cloud growth against the guided 57 to 63 percent range, capacity actually delivered, and any commentary on the credit outlook. Not the RPO headline.

Marvell is the cleanest business model of the three and the one I understand best. It ships product, books revenue, and gets paid, with no eight-year backlog requiring an act of faith. It is also priced for perfection at 81 times earnings after a 161 percent year, reporting the day after NVIDIA. If you want exposure to this theme without underwriting a single counterparty, this is the vehicle. If you want it at a sane entry, you probably want it after August 27, not before.

Cerebras is a venture position that happens to have a ticker. Beta above five, no profits, 68 times sales, one customer, and an accounting quirk that will manufacture optical misses through 2031. There is a genuine technology edge here and the AWS channel is real. But a stock that travels from $386 to $161 to $252 in ninety days is not an investment, it is a position you size in basis points and defend with rules you wrote before you entered.

The broader test arrives in the next three weeks. NVIDIA on August 26, Marvell on August 27, Broadcom shortly after, Oracle in the second week of September, and Marvell's investor day on October 6. That sequence will tell you more about whether the buildout is being financed or merely announced than any backlog figure will.

Fair winds and following seas.

---

Disclaimer: This post is for informational and educational purposes only and does not constitute investment advice, a recommendation, or an offer to buy or sell any security. Prices, market capitalizations, valuation multiples, and analyst targets are approximate and as of mid-August 2026, and will change. Price-to-sales figures are derived estimates, not reported metrics. Backlog attribution to specific customers reflects analyst estimates and press reporting rather than company disclosure in every case, and should be treated as an estimate. Forward-looking guidance is management's, not the author's. All three securities discussed are volatile, and one is a recent IPO with limited trading history and a reported beta above five. Verify all figures against primary filings before acting. Past performance does not guarantee future results. Consult a licensed financial professional regarding your specific situation.