Intel's Foundry Revenue Hit $5.8 Billion Last Quarter. 95% of It Was Intel Paying Itself.
Intel posted its strongest revenue growth in fifteen years on Thursday, beating estimates by $1.7 billion. The foundry business reported $5.8 billion in revenue. Subtract $5.5 billion in intersegment transfers and the external foundry generated $293 million, losing roughly $7 for every dollar an outside customer paid.
$293 million. Buried on page four of Intel's second-quarter earnings release, surrounded by no commentary, no year-over-year comparison, no analyst question on the ninety-minute conference call that followed, that single number captures something the $5.8 billion headline actively obscures: the amount of revenue Intel's foundry business collected from customers who are not Intel.
The headline number is $5.77 billion, which is what Intel Foundry reported as total quarterly revenue, up a respectable 31% year over year and comfortably beating the $5.55 billion that Wall Street's semiconductor analysts had penciled in as their consensus estimate for the segment. It sounds like a real business operating at real scale, the kind of number that justifies a 172% stock rally and tens of billions in government subsidies and a presidential photo op. But $5.5 billion of that total vanishes in a line item called "intersegment eliminations," which is the accounting entry for Intel's product divisions buying wafers from Intel's own manufacturing division, and when you strip those internal transfers away, what remains is a commercial foundry generating $293 million per quarter, roughly what GlobalFoundries makes in external revenue every eleven days.
The Turnaround Everyone Saw
Start with the numbers that matter to the stock price, because they are genuinely impressive and the market's twelve-percent after-hours pop is not irrational in the slightest. Second-quarter revenue hit $16.1 billion, a 25% leap from the same quarter last year that landed nearly $1.7 billion ahead of what analysts expected, making it the largest top-line surprise Intel has delivered since the dotcom era, the kind of print that forces even the bears to look up from their spreadsheets and acknowledge something has changed. Non-GAAP earnings per share landed at $0.42 against a consensus of $0.22. Almost double.
Data centers are the engine. Intel's Data Center and AI group posted $6.26 billion in revenue, a 59% year-over-year surge that beat estimates by nearly $900 million, driven almost entirely by a demand spike in CPUs for agentic AI inference workloads that Intel is uniquely positioned to absorb because it never stopped making x86 server chips while every other company in the industry spent three years chasing Nvidia's CUDA ecosystem. Irony compounding irony: Intel's inability to compete with Nvidia in GPUs left it as the default winner when the AI industry pivoted to inference and discovered it needed the processor architectures Intel had been quietly perfecting in obscurity.
Client computing pulled $8.88 billion against $7.89 billion expected. PC unit sales were actually down, but average selling prices climbed as Intel shifted its product mix toward higher-end silicon and abandoned the commodity market where margins evaporate, a strategic retreat that is paying off in dollars even as it concedes market share in units. Total non-GAAP operating income swung from a $503 million loss in Q2 2025 to a $2.77 billion profit in Q2 2026, an improvement of $3.27 billion in twelve months, which, to put an exclamation point on the scale of the turnaround, exceeds the entire annual operating profit Intel generated in 2023. Cash from operations: $7.0 billion. Gross margin expanded from 29.7% to 41.8%. Real results, all of them.
The Foundry Number Nobody Discussed
Now look at the foundry, because the foundry is the whole bet, the single strategic wager that separates Intel's turnaround story from a conventional semiconductor earnings beat and that justifies a valuation premium the CPU business alone would never command. When Lip-Bu Tan took the CEO job in March 2025, he inherited Pat Gelsinger's IDM 2.0 strategy: transform Intel from a company that designs and manufactures its own chips into a company that also manufactures chips for external customers, competing head-to-head with TSMC and Samsung for the foundry revenue that funds the next generation of process technology. The Trump administration bought in with a 10% equity stake and $8.5 billion in CHIPS Act subsidies, the largest government bet on a single semiconductor company in American history, backed by a national security argument that writes itself in three words: reduce Taiwan dependence.
Intel Foundry reported Q2 revenue of $5.77 billion. Impressive. But Intel's own financial statements show $5.5 billion in intersegment eliminations, the revenue Intel Foundry recorded from selling wafers to Intel's own product groups, and the external component, the revenue from customers who chose Intel Foundry over every alternative on the planet, was $293 million, a number so small relative to the headline that you could mistake it for rounding error if you were reading quickly, which is perhaps the point.
Run the math on that $293 million and the portrait sharpens into something uncomfortable. Intel Foundry's operating loss in Q2 was $2.1 billion, meaning the entire segment, internal and external revenue combined, lost $7.16 for every dollar an external customer paid. If you allocate the loss proportionally across total foundry revenue, the external $293 million is responsible for roughly $106 million in losses, but that proportional allocation is generous because external orders typically run on less-established process nodes where yields are lower and engineering support costs are higher and the actual per-wafer economics of serving an outside customer are almost certainly worse than the blended average suggests.
For comparison: TSMC's external foundry revenue in Q2 2026 was approximately $25.8 billion at a gross margin above 55%, Samsung Foundry reported roughly $4.6 billion in external revenue, GlobalFoundries posted $1.83 billion, and UMC pulled in $1.94 billion. Intel's $293 million places it below every major foundry and roughly on par with Tower Semiconductor, a specialty fab that focuses on analog, radio frequency, and power management chips and does not pretend to compete at the leading edge.
| Foundry | Q2 2026 External Revenue | Operating Margin |
|---|---|---|
| TSMC | ~$25.8B | ~47% |
| Samsung Foundry | ~$4.6B | ~5% |
| GlobalFoundries | $1.83B | ~23% |
| UMC | $1.94B | ~30% |
| Tower Semiconductor | ~$350M | ~15% |
| Intel Foundry (external) | $293M | -36% (blended) |
The Crossover Calculation
Intel Foundry's operating loss improved by $348 million quarter over quarter, from an implied $2.45 billion in Q1 to $2.1 billion in Q2. Zinsner attributed this to "higher yields, improved cycle times, and increased factory scale across Intel 43 and 18A." The cost trajectory on 18A is moving fast: the primary Panther Lake client CPU die cost dropped approximately 50% in the first half of 2026, with another 20% reduction targeted for the second half and what Intel described as "further meaningful reductions" planned for 2027.
If the $348 million quarterly improvement holds at a constant rate, Intel Foundry breaks even in six quarters, landing squarely in Q4 2027, a timeline that is suspiciously, almost theatrically convenient because it coincides almost exactly with the planned ramp of 14A risk production and the period when Tesla's "Terafab" AI chips would begin generating external revenue at volume and when the Apple deal, if it exists, would start contributing its first wafers.
Constant-rate improvement in semiconductor manufacturing is fiction, which anyone who has followed TSMC's node transitions or Samsung's 3nm struggles or Intel's own 10nm debacle already knows instinctively. Yields follow an S-curve: rapid improvement early in a node's life as the obvious defects get fixed, then a long plateau as the remaining issues become progressively harder to diagnose and exponentially more expensive to resolve, the kind of diminishing returns where the last five percentage points of yield cost more engineering hours than the first forty. Intel is currently riding the steep part of the 18A S-curve, which explains the 50% cost reduction in six months, but the next 50% will take longer and the 50% after that may never arrive at all.
A more realistic model assumes declining marginal improvement: if each quarter's gain is 80% of the previous quarter's, the sequence becomes $348M, $278M, $223M, $178M, $143M, $114M, a decelerating cascade that sums to $1.284 billion in cumulative improvement after six quarters instead of the $2.088 billion that the linear model produces. Under this scenario breakeven slides from Q4 2027 to Q2 2028, which is the exact quarter when 14A is supposed to enter high-volume manufacturing with Tesla as its anchor customer, turning what looked like a comfortably buffered schedule into a timeline balanced on a knife's edge where a single quarter of slippage on either side creates either a $500 million windfall or a $500 million hole.
What Tesla's "Terafab" Actually Means
Reuters confirmed that Tesla has committed to Intel's 14A process for what it calls the "Terafab" AI training chip. This is significant because Tesla is a real customer with real volume, not a paper design win. But the revenue math is more modest than the headlines suggest.
Automotive and AI training chips occupy different price bands than the smartphone and data center processors that drive TSMC's revenue. Tesla's AI training hardware competes with Nvidia's Blackwell architecture, which TSMC manufactures at scale on N4P and CoWoS advanced packaging. If Tesla's 14A commitment covers, say, 10,000 to 50,000 wafers per year at an estimated $15,000 to $20,000 per 14A wafer, the implied annual revenue is $150 million to $1 billion. At the midpoint, that roughly doubles Intel Foundry's external quarterly revenue. Meaningful, but not transformative.
Apple is the prize. President Trump announced in April 2026 that Apple had agreed to make processors with Intel, but neither company has confirmed a word of it. Apple currently buys approximately 20% of TSMC's leading-edge capacity. Securing even a fraction of Apple's volume at 14A would generate billions in annual external foundry revenue and validate Intel's process technology to every other potential customer. Without Apple or a customer of similar scale, the path from $293 million to foundry breakeven relies entirely on internal cost reduction.
The Samsung Precedent Nobody Wants to Discuss
Samsung Foundry followed a strikingly similar playbook. Samsung's mobile division provided guaranteed internal volume that funded the ramp of leading-edge nodes, while the foundry business pursued external customers including Qualcomm, Nvidia, and Google. By 2022, Samsung Foundry had captured roughly 12% of the global foundry market by revenue.
Then yields on Samsung's 3nm GAA process disappointed, and Qualcomm moved its Snapdragon flagship back to TSMC. Google's Tensor chips stayed with Samsung but remained a fraction of Google's total chip volume. Samsung Foundry's external market share has stagnated below 13% since 2023, despite billions in annual capital expenditure. The internal cross-subsidy sustains the business but has not generated the external momentum that would justify the investment on commercial terms alone.
Intel's foundry faces the same structural challenge: internal volume guarantees survival but does not prove competitiveness against the most reliable manufacturing operation in semiconductor history.
The CHIPS Act Subsidy Math
Intel has been awarded $8.5 billion in direct CHIPS Act subsidies and up to $11 billion in federal loans to support its domestic manufacturing buildout, making it by far the largest single recipient of semiconductor-specific industrial policy in U.S. history, a distinction that carries both financial leverage and political exposure in roughly equal measure. The stated purpose is to establish a U.S.-based alternative to TSMC at the leading edge.
But the subsidy creates a measurement problem that nobody in Washington seems interested in solving. If Intel Foundry's operating loss is shrinking by $348 million per quarter, some portion of that improvement necessarily comes from government subsidies flowing through the income statement as reduced depreciation on subsidized assets, lower-cost clean room construction, or workforce training credits, and Intel does not break out how much of the quarterly improvement reflects organic operational execution versus taxpayer-funded cost absorption. Without that decomposition, the question of whether Intel Foundry is becoming commercially competitive or merely becoming cheaper to subsidize has no empirical answer.
What We Don't Know
Intel's intersegment transfer pricing is opaque by design, and this opacity is doing more analytical work than any number in the earnings release. The $5.5 billion that Intel Products pays Intel Foundry is set internally, not negotiated at arm's length. Intel could, in theory, inflate foundry revenue by raising internal transfer prices, making the foundry look healthier while absorbing the cost in lower margins at the product level. The earnings release provides no mechanism for outsiders to verify whether the internal pricing reflects market rates. TSMC charges between $10,000 and $18,000 per N3 wafer depending on volume and complexity. Intel has not disclosed what it charges itself.
We also do not know Intel's foundry-specific capital expenditure, which is the single most important number for calculating return on invested capital and the single number Intel declines to provide. The company guided total 2026 net capex of $24 to $26 billion, meaningfully up from prior guidance, but did not break out the foundry's share. If the foundry consumes 60% to 70% of total capex, which is consistent with the scale of its manufacturing expansion, that is $14 to $18 billion per year in capital investment supporting a business that generates $293 million per quarter in external revenue. ROIC on that basis is not flattering.
What You Can Do
If you hold Intel stock, understand what you own. The 172% year-to-date gain prices in a successful CPU franchise, a growing data center business, and a foundry that becomes commercially viable. Two of those three are delivering. A $293-million-per-quarter business burning $2.1 billion, with breakeven contingent on yield curves, government subsidies, and an unconfirmed Apple deal. At $100, the stock prices in a lot of foundry optimism.
If you are a chip designer evaluating foundry options, Intel's 18A cost reductions suggest aggressive pricing is coming. A 50% die cost reduction in six months implies Intel is willing to buy market share at below-cost economics, subsidized by internal volume and government money. Get quotes now. The pricing will never be this favorable again.
If you are a policymaker overseeing CHIPS Act disbursements, demand the decomposition: how much of Intel Foundry's quarterly improvement is organic execution versus subsidy-funded cost reduction? The entire rationale for $8.5 billion in taxpayer investment is that Intel can become a commercially competitive foundry. A business generating $293 million in external revenue per quarter does not yet meet that test, and the public deserves to know whether the trajectory is real.
The Bottom Line
Intel's turnaround is real. Its CPU business is thriving on agentic AI demand, the data center group just posted 59% growth, and gross margins have expanded twelve points in a year. None of that is disputed. What is disputed, or should be, is the foundry. Intel Foundry reported $5.77 billion in revenue and $5.5 billion of it was Intel paying itself. External customers contributed $293 million. The operating loss was $2.1 billion. The breakeven math works if yield improvements continue at the current rate for six more quarters, if Samsung's precedent doesn't repeat, if Tesla's 14A volume materializes on schedule, and if Apple's rumored commitment becomes real. That is a lot of ifs for a business that has absorbed $8.5 billion in taxpayer subsidies and still loses $7 for every external dollar earned. The market is pricing in the answer. The data hasn't delivered it yet.