Is Elon Musk's Terafab Real, and Who Else Is Actually Competing With Nvidia?
From AMD to Google's TPUs to Musk's Terafab and Huawei: a guide to every real challenge to Nvidia's AI chip dominance, and what's actually investable.
So, Is Terafab Actually Real?
Yes. On March 21, 2026, Tesla and SpaceX announced "Terafab," a joint chip-manufacturing venture originally described as spanning design, fabrication, memory production, advanced packaging and testing all under one roof, with xAI (since folded into SpaceX, more on that below) as the third partner. On August 6, 2026, the companies confirmed the actual site: Grimes County, Texas, north of Houston, with a $16.8 billion first-phase investment, a facility described as more than 100 million square feet, and a promise of at least 3,000 jobs in the surrounding counties (TechCrunch; Electrek).
Notice the number. The March announcement talked about $25 billion. A May regulatory filing raised that to $55 billion for phase one, with a potential $119 billion across the full multi-phase buildout. The confirmed figure in August, once an actual site and initial investment were locked in, was $16.8 billion. That is not necessarily dishonest, early-stage megaprojects routinely reset their numbers as plans firm up, but it is a pattern worth tracking every time a new Terafab figure appears in a headline: which of the last four numbers is it actually citing?
Intel's role is real and specific: the company is supplying its 14A process node, its next-generation manufacturing technology that has not yet shipped commercially anywhere, reportedly offering 15% to 20% better performance and 30% higher density than Intel's current 18A node (Intel's own disclosed roadmap figures, as reported by Tom's Hardware and WCCFTech). Terafab would be 14A's first major external customer, a genuinely significant win for Intel's struggling foundry business regardless of how the rest of Terafab plays out.
A process node is simply the specific generation of manufacturing technology a factory uses to etch a chip design onto silicon, essentially a label for how small and how densely the transistors on that chip can be packed. Smaller, newer nodes fit more transistors into the same sliver of silicon, which generally means a faster, more power-efficient chip for the same size and cost. The industry used to name these generations after a physical measurement in nanometers, the "5nm" or "3nm" you may have seen elsewhere on this site, but Intel has since moved to naming its newest generations in Angstroms instead, where 18A and 14A no longer correspond to a literal, measurable distance so much as a generational label. What matters for us here is simpler: a lower number is a newer, more advanced generation, and 14A is genuinely one full generation ahead of anything shipping commercially from any chip factory in the world today, Intel's own included.
But read the filing, not just the press release. SpaceX's own S-1 reportedly describes the Terafab arrangement as a "general framework" with no binding commitments, no finalized intellectual property split between the partners, and no obligation for either side to keep participating (reported by techtimes). Separately, SpaceX did not discuss Terafab at all during its most recent earnings call, and Intel has reportedly been "cagey" when asked to specify exactly what it is contributing beyond the process node itself (TechCrunch). Electrek's coverage has been openly skeptical throughout, at one point describing the project as something that "reeks of desperation," and it is worth noting plainly that Musk has no prior semiconductor manufacturing experience.
What is it actually for? This is the detail most coverage buries. Musk has described the planned compute split as roughly 25% for Tesla's Optimus robots and about 75% for SpaceX's AI-driven spacecraft systems, not primarily for training xAI's Grok models the way most people assume when they hear "AI chip factory." The corporate structure behind this is itself new: SpaceX acquired xAI in an all-stock deal on February 2, 2026 (xAI valued at $250 billion, the combined entity at $1.25 trillion), and Musk dissolved xAI as a standalone company in May 2026, folding it into SpaceX as "SpaceXAI."
Terafab is a real, funded, genuinely ambitious project with a real semiconductor partner in Intel, not a hoax or vaporware. It is also, by its own regulatory filing's description, a loose framework rather than a locked-in commitment, its headline dollar figure has moved four times in five months, and its primary purpose appears to be securing captive compute for Musk's own robotics and aerospace ambitions rather than building a business that sells chips to the world. Treat every future Terafab headline with that context attached.
Nvidia Has Two Kinds of Rivals. Only One Can Actually Hurt It
Search "companies competing with Nvidia" and you get a list: AMD, Google, Amazon, Microsoft, Meta, Huawei, Cerebras, Terafab. Treating all of them as the same kind of threat is the single most common mistake in how this story gets covered, and it obscures the one distinction that actually matters for deciding what, if anything, is investable.
Built by one company for its own internal use. ASICsASIC (Application-Specific Integrated Circuit)A chip designed to do one job extremely well, rather than a general-purpose processor like a GPU. Google's TPU, Amazon's Trainium, Microsoft's Maia and Meta's MTIA are all AI ASICs, built by a hyperscaler for its own workloads and, unlike Nvidia's GPUs, generally not sold to outside customers.See all terms in the glossary designed for one company's specific workloads.
Generally not sold to outside customers. Competes with Nvidia for a share of that one company's own future budget, not for a sale in the open market.
Examples: Google TPU, AWS Trainium, Microsoft Maia, Meta MTIA, and Tesla/SpaceX's planned Terafab output.
You cannot buy the chip directly. You can only buy the parent company's stock, and the chip is one input among many into that company's overall business.
Designed to be sold to any buyer who wants an alternative to Nvidia.
Directly competes with Nvidia for a customer's purchase decision, the way a rival car maker competes for a car buyer.
Examples: AMD's Instinct line, Huawei's Ascend line, and independent chip startups like Cerebras and Groq.
Some of these are directly investable as their own equity, separate from any single AI lab or hyperscaler's fortunes.
Built by one company for its own internal use. ASICsASIC (Application-Specific Integrated Circuit)A chip designed to do one job extremely well, rather than a general-purpose processor like a GPU. Google's TPU, Amazon's Trainium, Microsoft's Maia and Meta's MTIA are all AI ASICs, built by a hyperscaler for its own workloads and, unlike Nvidia's GPUs, generally not sold to outside customers.See all terms in the glossary designed for one company's specific workloads.
Generally not sold to outside customers. Competes with Nvidia for a share of that one company's own future budget, not for a sale in the open market.
Examples: Google TPU, AWS Trainium, Microsoft Maia, Meta MTIA, and Tesla/SpaceX's planned Terafab output.
You cannot buy the chip directly. You can only buy the parent company's stock, and the chip is one input among many into that company's overall business.
Designed to be sold to any buyer who wants an alternative to Nvidia.
Directly competes with Nvidia for a customer's purchase decision, the way a rival car maker competes for a car buyer.
Examples: AMD's Instinct line, Huawei's Ascend line, and independent chip startups like Cerebras and Groq.
Some of these are directly investable as their own equity, separate from any single AI lab or hyperscaler's fortunes.
Nvidia can lose ground to a captive chip without losing a single sale to a competitor. If Google runs 80% of its own workloads on TPUs instead of Nvidia GPUs, that is real revenue Nvidia never collects, but Nvidia has not been beaten by a rival product in a bake-off. It has simply had one of its largest customers decide to become its own supplier for part of its needs.
AMD Is the Only One Actually Selling You a Way Out
AMD's Instinct line is the one product on this entire list that is a direct, drop-in-shaped alternative to an Nvidia GPU, sold to any buyer willing to switch. The MI350 (288GB of HBM3e memory, 8TB/s bandwidth) launched in Q3 2025, and AMD claims performance competitive with Nvidia's current generation. The MI400, expected in the second half of 2026, steps up to 432GB of HBM4 and 19.6TB/s of bandwidth, a 145% bandwidth increase over the MI350, positioned as a rack-scale competitor to Nvidia's own next-generation NVL144 systems.
Nvidia's overall AI GPU market share is reported around 85%, down from roughly 92% in 2024. AMD currently holds an estimated 5% to 7%, with some analysts projecting it could reach 12% to 15% by the end of 2026 as MI400 volume ramps.
AMD's ceiling here is not primarily about chip design quality. It is the same bottleneck this site covered in depth in our HBM and TSMC packaging article: Nvidia has reportedly reserved roughly 60% of TSMC's total 2026 advanced packaging capacity, leaving AMD, Broadcom, Google, Amazon and everyone else fighting over what remains. AMD can design the best GPU in the world and still be unable to ship enough of them, because the constraint sits one layer beneath the chip itself, in exactly the packaging chokepoint that article explained.
AMD is a real, growing, directly investable competitor (NASDAQ: AMD). It is not yet a peer to Nvidia by scale, and its growth trajectory is gated by the same physical bottleneck constraining everyone else on this list.
Google, Amazon, Microsoft and Meta Are All Quietly Doing the Same Thing
This is where most of the "competing with Nvidia" headlines actually live, and where the captive-versus-merchant distinction matters most.
Combined, hyperscaler custom silicon deployment is estimated at roughly 1.9 million accelerators in 2026, and custom ASIC shipments overall are reportedly growing around 44.6% year over year, nearly triple the growth rate of merchant GPUs. Perhaps the single most telling data point: Google has reportedly said more than three-quarters of the computation behind its Gemini models already runs on its own TPUs rather than Nvidia hardware.
Treat every "X% lower total cost of ownership" claim from a hyperscaler about its own chip exactly the way this site treats every other vendor-supplied efficiency number: real, but self-reported, not independently audited, and measured against a comparison point the vendor itself chose. Google's 44% TCO claim versus "a comparable Nvidia GB200 server" is Google's own comparison, not a third party's benchmark.
None of these four chips are meaningfully available for you or any outside company to buy and run in your own data center. That is precisely why they do not show up as a dent in Nvidia's merchant GPU revenue the way an AMD sale does, even as they quietly reduce how much of each hyperscaler's own future AI budget flows to Nvidia. The way to get exposure to any of this is to own the parent company: Alphabet, Amazon, Microsoft or Meta, our existing Alphabet analysis covers one of these four in full, not the chip project as a standalone investment, because there isn't one.
Okay, So What Is Musk Actually Building?
With the captive-versus-merchant framing established, Terafab's actual category becomes clear: it belongs firmly in the captive column, arguably the most captive project on this entire list. It is not designed to sell chips to outside buyers. Musk's own stated split, roughly a quarter of Terafab's output for Tesla's Optimus robots and three-quarters for SpaceX's AI-driven spacecraft systems, describes a facility built to solve one company group's own compute-supply problem, not to compete for Nvidia's customers.
What makes it structurally different from Google's, Amazon's, Microsoft's or Meta's captive chips is where it gets fabricated. Every one of those four hyperscaler chips is still made at TSMC, competing for the same scarce advanced packaging capacity AMD is fighting over. Terafab's entire premise, using Intel's 14A process instead, is a direct attempt to sidestep that exact bottleneck by building capacity outside the TSMC ecosystem altogether. If it works at anything like the scale described, it would be the first credible large-scale alternative to the TSMC chokepoint this site has spent an entire article explaining. If it doesn't, given the "general framework" language in SpaceX's own filing, it will have been an expensive, well-publicized bet that never became binding.
There is no honest way to call this one yet. Terafab is real money, a real semiconductor partner, and a real attempt to escape a real bottleneck. It is also, by SpaceX's own description, not yet a firm commitment, and its track record so far is four different headline dollar figures in five months.The Underdogs Betting the Chip Itself Has to Change
A third category exists alongside AMD's merchant GPUs and the hyperscalers' captive ASICs: standalone companies building AI chips with genuinely different architectures, not cheaper copies of an Nvidia GPU.
Cerebras builds wafer-scale chips, a single enormous chip covering nearly an entire silicon wafer rather than many small dies, a fundamentally different approach to the memory-bandwidth problem this site's HBM article covered in detail. Cerebras completed its IPO on May 14, 2026, closing its first day at a fully diluted valuation above $56 billion, reportedly the largest US tech IPO since Snowflake in 2020. Part of its pitch, per SemiAnalysis's Dylan Patel, is supply security: AI labs like OpenAI facing Nvidia allocation constraints and multi-month lead times want a credible second source, and Cerebras positions itself as exactly that.
Groq, whose LPU (language processing unit) architecture targets inference rather than training, took a genuinely stranger path, worth spelling out precisely because the headlines alone do not make it clear. On December 24, 2025, Nvidia agreed to acquire Groq's chip hardware, patents and founding leadership for roughly $20 billion, structured not as a normal buyout but as a non-exclusive technology licensing deal, effectively an acqui-hire: Groq's CEO Jonathan Ross, President Sunny Madra and key engineers moved to Nvidia along with the chip IP itself. Groq's separate cloud and inference-service business was explicitly carved out of that deal and continued on as its own independent company, which paid out roughly $7.6 billion (about $64 a share) to its shareholders from the proceeds, then re-staffed under new leadership and raised a fresh $650 million round on June 22, 2026 to build itself out as a standalone AI inference-cloud operator, now running 13 data centers across four continents and reportedly serving more than 5 million developers.
So, in plain terms: Groq's original chip technology and its founders now effectively sit inside Nvidia. The Groq name lives on as a separate, freshly re-funded cloud company with new leadership, not the same business that built the chip in the first place. It is not currently a simple public equity investment.
Smaller names in the same category, SambaNova, Tenstorrent (led by veteran chip architect Jim Keller) and Etched (which builds chips specialised only for transformer-model architecture), are real but thinner on public financial detail as of this writing, and none currently offer a straightforward retail investment route.
Cerebras is the one name in this entire article that is both architecturally distinctive and freshly, directly investable. That does not make it a good investment at a $56 billion valuation; it makes it a name worth understanding, not a name worth buying on the strength of a single paragraph. Apply the same discipline covered in our valuation guide before treating "this company might dent Nvidia" as a reason to buy anything.
China Didn't Beat Nvidia. Washington Did
This is the starkest number in the entire piece, and it comes directly from Nvidia's own CEO, not an analyst estimate. Jensen Huang has said publicly, in multiple venues through 2026, that Nvidia's AI chip market share in China went from roughly 95% to zero (reported by CNBC and multiple outlets covering his comments), adding: "I can't imagine any policymaker thinking that that's a good idea," and describing the outcome as having "already largely backfired" as a policy.
The mechanism is not competitive, it is regulatory. US export controls blocked Nvidia's most advanced AI chips from the Chinese market starting in late 2022, and successive rounds of restrictions closed off even China-specific compliant variants like the H20. Huawei's Ascend line filled the resulting vacuum. Its Ascend 950PR, introduced March 2026, is claimed by Huawei to deliver roughly 2.87 times an Nvidia H20's compute at FP4 precision, and analyst estimates put Huawei on track to control 50% to 60% of China's AI chip market by the end of 2026. ByteDance alone has reportedly committed $5.6 billion to purchase 750,000 Ascend units in 2026.
The genuinely open question, flagged by analysts covering this space including CSIS, is not whether Huawei has volume, it clearly does, but whether Huawei's CANN software stack becomes stable and developer-friendly enough at scale to become a real CUDA-equivalent moat of its own, or whether it remains a volume solution to a supply gap without the same lock-in Nvidia enjoys everywhere else. That distinction will determine whether this is a temporary, policy-created opportunity for domestic Chinese chipmakers or the birth of a second, durable AI compute ecosystem entirely separate from Nvidia's.
None of this is investable from outside China in any ordinary way. Huawei is privately held and under US sanctions. Cambricon, Alibaba's T-Head and Baidu's Kunlunxin are either unlisted internal divisions or thinly traded China A-share names without a simple access route for most international, including Indian, retail investors. This front matters enormously for understanding the global AI chip landscape and says nothing meaningful about what to put in a portfolio.
So... Is Nvidia Actually in Trouble?
Pull back and check the number that matters most: Nvidia's own reported results. In its Q1 FY2027 results (quarter ended April 26, 2026), Nvidia reported record revenue of $81.6 billion, up 85% year over year, with data center revenue of $75.2 billion, up 92%, and a gross margin near 75% (Nvidia's own newsroom). Guidance for the following quarter pointed to roughly $91 billion in revenue at a similar margin. That is not the financial signature of a company being meaningfully displaced.
What You Can Actually Buy (And What You Can't)
The Bottom Line
-
Terafab is real but looser than the headlines suggest. A confirmed $16.8 billion first-phase investment in Grimes County, Texas, with Intel supplying its next-generation 14A process, but described in SpaceX's own filing as a general framework without binding commitments, and its stated purpose is powering Tesla's Optimus robots and SpaceX's spacecraft systems, not selling chips commercially.
-
The single most useful lens for this entire story is captive versus merchant. Google, Amazon, Microsoft, Meta and now Terafab are building chips to buy less Nvidia for themselves, not to sell chips that compete with Nvidia in the open market. Only AMD, Huawei and independent startups like Cerebras actually do the latter.
-
AMD is the one real, growing, directly investable merchant rival, at an estimated 5% to 7% of the AI GPU market today, but its growth is capped by the same TSMC advanced packaging bottleneck covered in our HBM article, since Nvidia has reportedly reserved roughly 60% of that capacity.
-
China is a policy story, not a competitive one. Nvidia's own CEO says its China market share went from 95% to zero because of export controls, not because Huawei out-competed it in an open market, and none of the beneficiaries (Huawei, Cambricon, Alibaba's T-Head, Baidu's Kunlunxin) are accessible to ordinary international investors.
-
Cerebras is the one independent challenger you can actually buy, freshly public at a $56 billion-plus valuation after its May 2026 IPO, building architecturally distinct wafer-scale chips rather than a cheaper GPU copy, though a large valuation is itself a reason for caution, not excitement.
-
None of this shows up yet in Nvidia's own numbers. Q1 FY2027 revenue of $81.6 billion, up 85% year over year, with 75% gross margins, is not the financial profile of a company losing its core business, even as real, measurable erosion is happening at the edges: China entirely, hyperscaler internal workloads partially, AMD's merchant share slowly.
-
The CUDA moat, not any single competitor, remains the real story. Every captive and merchant challenger covered here still has to solve the same problem Nvidia solved starting in 2006: building a software ecosystem developers won't want to leave, covered in full in our Nvidia business analysis.
Where All These Numbers Came From
- TechCrunch, "Tesla and SpaceX will invest $16.8B to start building 'Terafab' chip factory in Texas," August 6, 2026. Confirmed site, first-phase investment figure, and Intel's contribution.
- Electrek, "Tesla, SpaceX confirm 'Terafab' chip fab site," August 6, 2026, and Electrek's earlier March 2026 coverage. Skeptical framing and cost-figure history.
- techtimes, "SpaceX Confirms Terafab in Texas; Own Filing Shows $38B Gap in First-Phase Figures," August 2026. The S-1 "general framework" characterisation.
- Tom's Hardware, on Terafab's use of Intel's 14A process. Process node details and Intel's roadmap figures.
- Yahoo Finance, "Jensen Huang says Nvidia went from 95% market share in China to 0%." Direct quotes from Nvidia's own CEO.
- NVIDIA Newsroom, official Q1 FY2027 financial results. Revenue, data center revenue, and gross margin figures, from the company itself.
- Hyperscaler custom silicon figures (Google TPU, AWS Trainium, Microsoft Maia, Meta MTIA) as reported across 2026 industry coverage including Tom's Hardware's custom AI ASIC tracking and Introl's Custom Silicon Inflection 2026 research.
- AMD Instinct MI350/MI400 specifications and market share estimates, and TSMC CoWoS capacity allocation, as reported across 2026 semiconductor trade press, cross-referenced against this site's own HBM and TSMC packaging article.
- Cerebras IPO valuation and positioning, as reported by Morningstar and SemiAnalysis (via chipstrat.com). Nvidia's December 2025 licensing deal for Groq's chip business and founding team, and Groq's own June 2026 funding round as an independent cloud company, as reported by CNBC and TechCrunch.
- Huawei Ascend chip specifications, China market share projections, and ByteDance's purchase commitment, as reported across 2026 coverage including Tom's Hardware and FourWeekMBA, with CSIS providing the software-moat framing.
Disclaimer
Nothing on this site is investment advice. All content is for educational and informational purposes only. Do your own research and consult a registered financial adviser before making any investment decisions.
Finished reading? Mark this article to track your learning progress.
Software Engineer, Self-Taught Investor
Software engineer who started learning about money in 2016 after a layoff coincided with a new home loan. Went from bank deposits to mutual funds to picking stocks in India and the US, learning through YouTube, screener.in, TradingView, and the hard way. Still learning. This site is her notes made public — for education and sharing only, not financial advice.
