Alphabet (Google) Places Order with Intel for 3 Million TPU Chips
According to reports from The Information, Google has ordered the production of over 3 million of its own Tensor Processing Units (TPUs) at Intel's facilities in 2028. This decision aims to diversify the supply chain away from TSMC and strengthen Intel's position in contract chip manufacturing.
Analysis: Google Orders 3 Million TPUs from Intel — End of TSMC's Monopoly or Beginning of the End for Intel Foundry?
When "Plan B" becomes a strategy, it means "Plan A" is already cracking at the seams.
On June 7, 2026, Reuters, citing a report from The Information, broke news that made analysts on Wall Street choke on their coffee: Google placed an order for the production of over 3 million of its own TPUs (Tensor Processing Units) at Intel's facilities. Delivery date: 2028. Intel's stock surged 13% within hours. The news was picked up by Yahoo Finance, Quartz, and The Next Web.
At first glance, it's a perfect story: an American tech giant saves an American chipmaker, diversifies supply chains away from Taiwan's TSMC, and strengthens national security. A beautiful picture for the White House, the Pentagon, and Intel's shareholders.
But let me tell you why this "victory" smells like kerosene. Why Google's order is not so much a triumph for Intel as a cry of desperation from the entire AI chip market. And why, in 90 days, we may see the flip side of this coin — when Intel fails to deliver on its promises, and TSMC finally crushes competitors with price and technology.
[The Core]: What's Really Happening
Official version: Google has ordered TPU production from Intel for 2028. Why 2028? Because by then, Intel promises to perfect its 14A process and ramp up sufficient capacity. Unofficial version, which you won't hear in press releases: Google is panicking. TSMC, which currently produces the vast majority of advanced AI chips (including Google's current TPUs on the N3 process), cannot keep up with demand.
What does "cannot keep up" mean? TSMC has already announced an increase in capital expenditure from $52 billion to $56 billion in 2026, but that's still not enough. TSMC's Senior Vice President K.K. Ho recently stated that the company is "ready for growth thanks to new business opportunities," but that's a diplomatic way of saying, "We can't produce everything they want from us."
The problem is compounded by the fact that the bottleneck is not so much the silicon wafers themselves, but advanced packaging. TSMC's CoWoS technology, necessary for connecting chips with HBM memory, is in short supply. This is where Intel has a temporary advantage with its Foveros and EMIB technologies, which Google tested for several months before making the decision.
The essence of the deal: Google is not abandoning TSMC. By no means. TSMC will remain the primary TPU manufacturer at least until 2027-2028. But Google is hedging its bets. 3 million chips from Intel is a "Plan B" that has grown so large it has become a standalone strategy. According to Morgan Stanley estimates, Google's total TPU production in 2027-2028 will exceed 6 million units, so Intel's share will be roughly 30-40%.
Timeline and Context
To understand why this deal became possible, we need to trace the chain of events over the past 18 months. This is not a story of a single decision, but a series of crises that forced Google to act.
Early 2025: TSMC informs clients that its capacity on advanced nodes (7nm and above) is "roughly three times lower than expected demand from key clients." NVIDIA, AMD, Apple, Qualcomm — all are lining up. Google, which is not TSMC's largest client (unlike NVIDIA, which takes the lion's share), realizes that in any shortage, its TPUs will be the first to suffer.
April 2025: Google announces the seventh generation TPU — Ironwood. A chip that, in FP8 performance (4600 TFLOPS), slightly edges out NVIDIA's B200 (4500 TFLOPS), while consuming only 157W versus 700W for the competitor. This is a breakthrough. But Ironwood is manufactured on TSMC's N3 process. Google finds itself trapped by its own success: the more clients connect to TPUs (Anthropic, all 17 national laboratories of the US Department of Energy, Citadel Securities), the more acute the capacity problem becomes.
February 2026: Meta signs a multi-year agreement for access to Google's TPUs. Meta represents a massive volume of computing. If previously TPUs were used mainly inside Google and by a few select partners, now mass commercial deployment begins. TSMC physically cannot ramp up capacity fast enough.
April 2026: Intel CEO Lip-Bu Tan, at the Morgan Stanley Technology, Media & Telecom Conference, states that demand for the 18A process for Panther Lake exceeds Intel's capacity. This is a strange statement — Intel cannot even meet its own demand, yet it takes on Google? The paradox resolves when it becomes clear that the focus is primarily on packaging, not full-cycle production.
June 2026 (now): News of Google's order. But in parallel — and the media barely noticed — NVIDIA is also testing Intel's 18A for its future chips, including the Feynman processor for 2028, which is supposed to combine four graphics chips into one. If NVIDIA follows Google's example, it will be a tectonic shift.
Who Wins and Who Loses
Winner #1: Intel (stock and market confidence). On this single news, Intel's market cap grew by tens of billions of dollars. This is not just money — it's trust. For the past three years, Intel Foundry Services (IFS) has been generating multi-billion dollar losses. The Google contract is the first real proof that IFS can attract hyperscaler-level clients. Additionally, Intel already has contracts with Microsoft to produce Maia 2 chips on the 18A process, and with Tesla for the Terafab factory in Austin.
Winner #2: Google (strategically). Google gains leverage over TSMC. When your supplier knows you have an alternative, they start offering better terms. But more importantly, Google reduces geopolitical risk. The US is increasingly concerned about the concentration of advanced chip manufacturing in Taiwan. The contract with Intel is "American manufacturing for an American company," which fits perfectly into the CHIPS Act narrative.
Winner #3: NVIDIA (paradoxically). NVIDIA did not directly order from Intel, but it is actively testing 18A. Why is this good for NVIDIA? Because if Intel can relieve TSMC, TSMC can allocate more capacity to NVIDIA. Moreover, NVIDIA has already invested in Intel through a $7 billion private placement. Intel's growth benefits NVIDIA as a shareholder.
Loser #1: TSMC (long-term). TSMC remains king, but its monopoly has cracked. Clients have realized that keeping all eggs in one basket is risky. TSMC is trying to expand in the US — the first fab in Arizona is already built (mass production started in Q4 2024), the second fab is completed and will start 3nm chip production in the second half of 2027, a third is under construction, and approval for a fourth has been obtained. But this won't solve the problem instantly. Institutional investors forecast that by 2030, TSMC's capacity ratio between Taiwan and the US will be 70% to 30%. That means 30% of capacity will be on territory not directly under threat from China. But 70% remains vulnerable.
Loser #2: Samsung Foundry. Samsung is also trying to become an alternative to TSMC. But Google chose Intel, not Samsung. This is a blow to the Korean giant's ambitions, especially painful given that Samsung just received an order from Tesla. Now the battle for second source is between Intel and Samsung, and Intel just scored an important goal.
Loser #3: Broadcom. Broadcom has been Google's partner in TPU development for decades. Google designs the chips, Broadcom helps with integration and manufacturing. With part of the orders moving to Intel, Broadcom's role may diminish. Meta is also considering using TPUs, further distancing Meta from Broadcom.
What the Media Isn't Telling You
The media narrative — "Intel is saved, Google made the right choice" — is not just a simplification but outright manipulation. Let's look behind the scenes.
Insight #1: Intel 18A is not TSMC N3P. Not even N2.
Let's face reality. Intel's 18A process, which is supposed to be the basis for Google's chip production, still has yield issues. Intel CFO David Zinsner acknowledged that the 18A process shows "some variability in yield" — some wafers come out with very low yields. And although he says "optimization is ahead of schedule," Intel's historical track record is full of examples where "ahead of schedule" turned into "two years late."
Moreover, Google ordered production for 2028. Why not 2026 or 2027? Because Intel simply isn't ready earlier. This is an "installment order" — Google believes that by 2028, Intel will solve its problems. But if it doesn't? The contract likely includes penalties, but they won't compensate Google for lost time and capacity.
Insight #2: This is a packaging contract, not a wafer production contract.
Notice the wording in the original reports from The Information: the order was preceded by "months of testing Intel's chip packaging technology." This is key. TSMC's bottleneck is not so much the chips themselves, but CoWoS packaging. Intel has alternative technologies — Foveros and EMIB. Google will likely produce the actual TPU dies at TSMC (because quality is higher and yields better), and then send them to Intel for packaging.
This is not "full production" at Intel. It's "partial production." Google retains the ability to return to TSMC for the full cycle at any time if Intel fails. But for PR and investors, this deal is sold as "Google chooses Intel." Nice, but not accurate.
Insight #3: The role of the US government.
Behind this deal is the invisible hand of Washington. The CHIPS Act provided Intel with $7.9 billion to expand US production. But this money comes with conditions: Intel must show real results and attract major clients. Google is the ideal candidate. I'm almost certain that behind-the-scenes "briefings" took place at the White House with Alphabet's leadership. "National security," "technological sovereignty" — these words were uttered more than once.
Furthermore, the US Department of Defense is interested in having critical AI chips produced on American soil, not in Taiwan, which could be blockaded by China. Google, which has contracts with the Pentagon and national laboratories, came under pressure: diversify production, or we'll find another cloud provider. This is not conspiracy theory; it's standard practice with federal contractors.
Forecast: Next 30 Days and 90 Days
Next 30 Days (until mid-July 2026)
1. Official confirmation from NVIDIA. NVIDIA is currently testing Intel 18A and the technology for combining four GPUs into one (for the Feynman 2028 architecture). Within a month, NVIDIA will make an announcement — either expanding testing or placing a preliminary order. I'm betting on "expanded testing" without firm commitments. NVIDIA is too smart to bet on Intel prematurely.
2. Correction of Intel's stock. 13% growth in one day is overheating. When emotions settle and analysts start asking uncomfortable questions about yields and real timelines, the stock will correct by 5-7%. This is not panic; it's a "return to reality."
3. Statements from TSMC. TSMC will have to respond. Most likely, they will announce accelerated capacity expansion in Arizona or a new round of investment. Expect news that a fifth fab in Arizona is already in the plans, and TSMC's total US investment could reach $250 billion.
Medium-Term Forecast (90 days, until September 2026)
1. Meta and Amazon intensify search for TSMC alternatives. Google set an example. Meta already has access to TPUs but does not produce its chips at Intel. Amazon has its own Trainium and Inferentia chips. Both companies will start negotiations with Intel for trial batches. If Intel can close another hyperscaler-level deal by the end of 2026, it will change the entire market dynamic.
2. Broadcom begins to lose ground. Broadcom's stock could correct by 8-10% if investors decide that the era of custom ASICs is ending, replaced by Intel's "integrated production + packaging." This would be an overreaction, but markets are often excessive. In reality, Broadcom will remain an important partner for Google for future TPU generations — but investors won't understand that immediately.
3. Growing interest in Intel from the automotive industry. Intel already has a contract with Tesla. After the Google news, automakers (especially German ones — Volkswagen, Mercedes) will start considering Intel as an alternative to Qualcomm and NVIDIA for autonomous driving systems. Intel could sign 2-3 contracts with automakers by the end of 2026 for ADAS chip production.
4. Main risk: a new round of US sanctions against China. The Biden administration or its successor may impose new restrictions on the export of advanced chips to China. This would hit TSMC (Chinese clients, albeit small) and indirectly help Intel, as production on American soil is not subject to export restrictions as strictly. Watch for news from Washington — they could be a catalyst for Intel.
Summary for those who read to the end: The Google-Intel contract is a historic deal, but not for the reasons the headlines suggest. This is not "Intel is back," but "Google hedged against disaster." Intel's real victory will come not when it receives the order, but when it can fulfill it. 2028 is an eternity in the chip world. By then, TSMC will be at 1.6nm (A16) or even 1.4nm, while Intel — if lucky — will be at 14A. The technology gap may not shrink; it may widen.
But as a financial bet and a geopolitical maneuver, the deal is flawless. Intel got oxygen. Google got insurance. TSMC got a warning. And we got the most exciting showdown in the semiconductor industry in the last 20 years. Buckle up — it's going to be a bumpy ride.
— Editorial Team
No comments yet.