Microsoft is preparing to show the world its most important piece of silicon since it entered the AI chip race. According to Reuters, citing a report from The Information published on August 10, 2026, Microsoft plans to unveil its Maia 300 AI accelerator this fall, with some reporting suggesting a reveal as early as September. The chip is meant to do something Microsoft has struggled with for two years: give Azure a real answer to Nvidia’s Blackwell GPUs without leaning on a single supplier for the compute that now underpins Copilot, Azure OpenAI Service, and nearly every AI product Microsoft ships.
This is the latest flashpoint in the broader AI hardware race. The timing matters, and Microsoft has already lost ground on its own custom-silicon roadmap once. Reuters reported back on June 27, 2025, also citing The Information, that the chip now known as Maia 300 had its mass production pushed back by at least six months, sliding a planned 2025 ramp into 2026. A year later, the company is trying again, and this time the numbers being discussed with its foundry partner are large enough to signal Microsoft wants Maia to matter at scale, not just as a symbolic in-house project.
What Microsoft Has Actually Confirmed So Far
Strip away the speculation and the confirmed record is fairly narrow. Reuters and The Information have reported that Microsoft intends to unveil Maia 300 in fall 2026. Separately, RuntimeWire and TrendForce both reported on August 10 and 11 that Microsoft is in discussions with TSMC to secure manufacturing capacity for more than 300,000 Maia 300 chips for delivery in 2027. That figure, repeated independently by two outlets citing the same underlying report chain, is the clearest signal yet that Microsoft wants Maia 300 to move past pilot-scale deployment.
Everything below that headline, the transistor counts, the memory figures, the exact process node, comes from industry reporting and leaks rather than a Microsoft datasheet. TechTimes published detailed specifications on August 11 describing a chip built on TSMC’s 3-nanometer process with more than 140 billion transistors. Digiopedia, in a piece the same week, repeated the 3nm claim but flagged that some semiconductor-industry sources have also floated a 2nm node, while explicitly stating that Microsoft has not officially confirmed the process, the HBM generation, or the bandwidth figures circulating online. That caveat is worth holding onto as you read the rest of this piece.
Inside Maia 300: The Specs Reported So Far
According to TechTimes, the leaked configuration includes 216 GB of HBM3e memory, 7 TB/s of memory bandwidth, and 272 MB of on-chip SRAM, a cache figure the outlet says tops the 192 MB on Nvidia’s Blackwell die. Digiopedia adds that the chip is expected to support native FP8 and FP4 tensor computation, the same low-precision math formats that Nvidia’s Blackwell and Google’s newest TPUs lean on for large-scale inference and training throughput.
Confirmed vs. Still Unverified
None of those memory or SRAM figures have been confirmed by Microsoft directly. Digiopedia’s reporting draws a clear line: process node, HBM capacity, and bandwidth are all industry estimates built from supply-chain sourcing, not disclosures from Redmond. Readers comparing Maia 300 against Nvidia or Google’s shipping products should treat the Maia numbers as directional rather than final until Microsoft’s own event happens.
The TSMC Bottleneck Nvidia Already Controls
Even if Maia 300’s specs land close to what’s been reported, Microsoft still has to get the chips built. TechTimes reports that Nvidia controls roughly 60% of TSMC’s advanced packaging capacity, the same CoWoS-style packaging that every high-bandwidth-memory AI chip depends on. That single data point explains a lot about why Microsoft’s 2025 ramp slipped and why the 300,000-unit 2027 target, reported by both RuntimeWire and TrendForce, is being negotiated now rather than promised outright. Packaging capacity, not wafer starts, has become the real chokepoint in the AI chip supply chain, and Microsoft is competing for it against the same customer it’s trying to displace inside its own datacenters.
Our earlier coverage of Nvidia’s HBM4 memory yield gains for the Rubin platform covered the same packaging math from Nvidia’s side of the table. Maia 300 is a reminder that every hyperscaler chasing custom silicon is fighting over the same limited packaging queue at TSMC.
Maia 300 vs. Nvidia Blackwell: How the Numbers Stack Up
Nvidia’s Blackwell lineup is the target Microsoft is measuring itself against, even if the company won’t say so publicly. The Register reports that Nvidia’s B200 delivers roughly 4.5 PFLOPS of dense FP8 compute per chip, while the higher-end GB200 and GB300 configurations push closer to 5 PFLOPS FP8. On the memory side, AI-chip tracking sites list the Blackwell Ultra B300 at 288 GB of HBM3e with about 8 TB/s of bandwidth, and note pricing in the neighborhood of $40,000 to $50,000 per accelerator, though that figure comes from third-party trackers rather than an official Nvidia price list, since Nvidia does not publish list prices for its data-center parts.
Set against that, Maia 300’s reported 216 GB of HBM3e and 7 TB/s of bandwidth land below the B300 on both counts, but ahead of the B200’s 192 GB memory pool. In other words, if the leaked numbers hold, Microsoft appears to be building something that slots between Nvidia’s mainstream and flagship data-center parts rather than trying to beat the B300 outright. That’s a pragmatic target: Microsoft doesn’t need Maia 300 to be the fastest chip in the world, it needs it to be good enough for a meaningful share of Azure’s own inference workloads at a lower internal cost than buying more Nvidia silicon.
How Google’s TPU v7 Ironwood Already Got There First
Google isn’t waiting for Microsoft to catch up. The Register and AI Compute Tracker both report that Google’s TPU v7, code-named Ironwood, reached general availability on Google Cloud in March 2026, delivering about 4.6 PFLOPS of FP8 compute, 192 GB of HBM3e, and roughly 7.37 TB/s of memory bandwidth. That bandwidth figure actually edges out Maia 300’s reported 7 TB/s, even though Maia’s memory capacity is slightly larger on paper. AI Compute Tracker also reports that Anthropic has committed to as many as 1 million Ironwood chips, a scale commitment that shows just how far ahead Google’s custom silicon program already is in terms of real customer adoption rather than announced intent.
TrendForce’s own assessment, published alongside its Maia 300 coverage, put it plainly: Google and AWS are already shipping their newest custom AI accelerators at scale, while Microsoft is still working to unveil and then ramp its next-generation chip. Microsoft isn’t first to this race, and by the numbers currently public, it isn’t fastest either.
AWS Trainium 3 Sets the Bar for Rack-Scale Delivery
Amazon’s custom silicon program offers a third comparison point. NerdLevelTech and CloudRPS report that AWS Trainium 3 delivers approximately 2.52 PFLOPS of FP8 compute per chip, paired with 144 GB of HBM3e and 4.9 TB/s of bandwidth, and reached general availability in December 2025. AWS packages the chip into Trn3 UltraServers, a rack-scale system built to compete directly with Nvidia’s NVL72 racks rather than chip-for-chip. AI Compute Tracker puts Trainium 3’s real-world throughput at roughly 4.4 times that of the previous Trainium 2 generation at similar power draw, a generational jump that came from architecture and packaging improvements rather than a simple die shrink.
Against Trainium 3, Maia 300’s reported specs look stronger on paper: more memory, more bandwidth, and presumably more raw FP8 throughput given the larger reported SRAM pool. But AWS already has Trainium 3 in production racks serving customers today, while Maia 300 remains an unveiling that hasn’t happened yet.
Where AMD’s MI400 Fits Into the Picture
AMD occupies a different lane entirely. Industry commentary tracked by outlets like AI Compute Tracker and NerdLevelTech frames AMD’s MI400 series as the main merchant-silicon alternative to Nvidia, the chip cloud providers and enterprises buy off the shelf rather than design in-house. Maia 300, by contrast, is part of the hyperscaler custom-chip cohort built specifically to reduce a single company’s dependence on both Nvidia and AMD. The two products aren’t really competing for the same customer: MI400 sells to anyone who writes a purchase order, while Maia 300 exists to keep more of Azure’s own AI spending inside Microsoft’s balance sheet. Our recent look at Nvidia’s biggest customers building their own rival chips covers this same dynamic playing out across five major cloud providers at once.
Maia 200: What’s Already Running in Azure Today
Maia 300 doesn’t arrive in a vacuum. NerdLevelTech reports that Maia 200, the current generation, already runs on TSMC’s 3nm process with more than 140 billion transistors, 216 GB of HBM3e, and 7 TB/s of bandwidth, and is deployed in Azure today for inference workloads. The overlap between Maia 200’s confirmed specs and Maia 300’s leaked specs is striking. It suggests Maia 300 is closer to a refined, higher-efficiency evolution of an already-shipping design than a ground-up architectural rewrite, an interpretation that lines up with how Microsoft has approached custom silicon since it first started deploying Maia parts internally.
The Bigger Pattern: Every Hyperscaler Now Designs Its Own Chips
Microsoft’s push into custom silicon didn’t happen in isolation. Google has spent the better part of a decade on TPUs, AWS has iterated through three generations of Trainium and multiple generations of Inferentia, and Meta has its own MTIA accelerators for internal workloads. What changed over the past 18 months is scale and urgency. Nvidia posted a record $96.2 billion quarter, and every cloud provider buying that much Nvidia silicon has a direct financial incentive to build an internal alternative, even a partial one. OpenAI is reportedly working on its own accelerator too, something we detailed in our reporting on the so-called Jalapeño chip aimed at Nvidia’s data-center margins.
Custom silicon doesn’t need to beat Nvidia on every workload to matter financially. It just needs to shift enough inference traffic away from purchased GPUs to bend a company’s own AI infrastructure cost curve. That’s the actual business case behind Maia 300, Ironwood, and Trainium 3 alike, and it’s why TrendForce’s framing of Microsoft as playing catch-up doesn’t mean Maia 300 is pointless. It means Microsoft is running the same playbook as its rivals, just a generation or two behind on execution.
Market Impact: What This Means for Azure, Nvidia, and AI Pricing
For Azure customers, a successful Maia 300 ramp could eventually mean cheaper inference pricing on Microsoft’s own AI services, since Microsoft wouldn’t need to pass along Nvidia’s margins on every token processed through Copilot or Azure OpenAI Service. For Nvidia, the risk isn’t that Maia 300 replaces Blackwell wholesale, it’s that Microsoft buys fewer incremental GPUs for the specific inference workloads Maia 300 can handle, while still buying Blackwell and eventually Rubin-generation parts for frontier model training. Our coverage of AWS’s own 2-million-GPU deal with Nvidia shows the same dynamic: even hyperscalers building custom chips keep buying enormous volumes of Nvidia hardware at the same time.
None of the reporting reviewed for this piece includes a stock-price reaction specific to the Maia 300 news. Reuters’ coverage placed the story in the broader AI infrastructure context without citing a market move tied directly to it, which fits the pattern for a chip that has been reported on through leaks and planning documents rather than a formal product launch with financial guidance attached.
Head-to-Head: The Five Chips Defining the 2026-2027 AI Hardware Race
| Accelerator | Maker | Status (Sept 2026) | Memory | Bandwidth | Reported FP8 Compute |
|---|---|---|---|---|---|
| Maia 300 | Microsoft | Unveiling planned fall 2026 (unconfirmed specs) | 216 GB HBM3e* | 7 TB/s* | Not disclosed |
| Maia 200 | Microsoft | Deployed in Azure (inference) | 216 GB HBM3e | 7 TB/s | Not disclosed |
| Blackwell B200 | Nvidia | Shipping | 192 GB HBM3e | Not stated | ~4.5 PFLOPS |
| Blackwell Ultra B300 | Nvidia | Shipping | 288 GB HBM3e | ~8 TB/s | ~5 PFLOPS (GB300-class) |
| TPU v7 “Ironwood” | GA since March 2026 | 192 GB HBM3e | ~7.37 TB/s | ~4.6 PFLOPS | |
| Trainium 3 | Amazon | GA since December 2025 | 144 GB HBM3e | 4.9 TB/s | ~2.52 PFLOPS |
Timeline: How the Custom Silicon Race Reached This Point
| Date | Event | Source |
|---|---|---|
| June 27, 2025 | Microsoft’s next-gen Maia chip mass production delayed roughly six months, sliding into 2026 | Reuters, citing The Information |
| December 2025 | AWS Trainium 3 reaches general availability via Trn3 UltraServers | AI Compute Tracker |
| March 2026 | Google TPU v7 “Ironwood” reaches general availability on Google Cloud | The Register, AI Compute Tracker |
| August 10, 2026 | Microsoft reportedly plans to unveil Maia 300 this fall, possibly September | Reuters, citing The Information |
| August 10-11, 2026 | Microsoft in talks with TSMC for capacity to build 300,000+ Maia 300 chips for 2027 | RuntimeWire, TrendForce |
| August 11, 2026 | Leaked specs describe TSMC 3nm process, 140B+ transistors, 216 GB HBM3e | TechTimes |
| 2027 (target) | Planned volume delivery window for Maia 300 at scale | RuntimeWire, TrendForce |
Historical Context: Microsoft Is Playing Catch-Up on Its Own Silicon
Microsoft entered the custom AI silicon race later than most of its peers and has already slipped its own internal schedule once, per Reuters’ June 2025 reporting. Google’s TPU program stretches back roughly a decade for internal use before opening up to cloud customers. Amazon has been iterating on Inferentia and Trainium chips since 2018. Against that backdrop, Maia 300 represents Microsoft’s attempt to close a multi-year gap in a single generational jump, backed by a TSMC capacity request, per RuntimeWire and TrendForce, large enough to signal real intent rather than another internal pilot program.
Why the Six-Month Delay Still Matters
The 2025 slip Reuters reported wasn’t just a scheduling footnote. It pushed Microsoft’s ramp directly into a window when Google and AWS were both bringing newer-generation chips to general availability, Ironwood in March 2026 and Trainium 3 in December 2025. Every month Maia 300 slips is another month Azure keeps buying more Nvidia GPUs to cover the gap, which is expensive at a moment when Nvidia’s own data-center margins remain among the highest in the semiconductor industry.
What Happens Next: Five Things to Watch
- Expect Microsoft to keep Maia 300 as an internal-only chip for Azure’s own AI services rather than selling it externally, following the same closed strategy Google used with TPUs for years before opening limited external access.
- TSMC’s advanced packaging queue, with Nvidia controlling roughly 60% of it per TechTimes, will likely remain the binding constraint on how fast Microsoft can actually hit the reported 300,000-unit 2027 target.
- Microsoft’s fall unveiling should resolve the process-node dispute between the 3nm and 2nm claims currently circulating, since Digiopedia notes neither has been officially confirmed.
- Watch for whether Microsoft discloses any FP8/FP4 throughput numbers at the unveiling. Right now Maia 300 is the only chip in this comparison with no reported compute figure at all, which makes it hard to judge against Blackwell, Ironwood, or Trainium 3 on performance rather than just memory specs.
- Don’t expect Microsoft to publish per-chip pricing. Unlike Nvidia’s GPU ecosystem, where third-party trackers estimate list prices even without official numbers, Maia’s economics have stayed entirely internal to Azure’s cost structure so far, and there’s no indication that changes with Maia 300.
Frequently Asked Questions
What is the Microsoft Maia 300?
Maia 300 is Microsoft’s next-generation custom AI accelerator chip, built for Azure’s AI infrastructure. Reuters, citing The Information, reports Microsoft plans to unveil it in fall 2026, with production targeting more than 300,000 units for 2027 delivery, according to RuntimeWire and TrendForce.
When will Maia 300 launch?
Microsoft is reportedly planning to unveil Maia 300 in fall 2026, with some reports suggesting a reveal as early as September, per Reuters and The Information. Volume production for 2027 is being negotiated with TSMC, according to RuntimeWire and TrendForce, so widespread availability likely won’t arrive until sometime in 2027.
How does Maia 300 compare to Nvidia’s B200 and B300?
Based on leaked specs reported by TechTimes, Maia 300’s 216 GB of HBM3e memory sits above Nvidia’s B200 (192 GB) but below the Blackwell Ultra B300 (288 GB, per AI-chip trackers). No FP8 compute figure has been reported for Maia 300, so a direct performance comparison isn’t yet possible, and Microsoft has not confirmed the memory numbers itself.
Is Maia 300 faster than Google’s TPU v7 Ironwood?
It’s not possible to say yet. Ironwood is already in general availability since March 2026 with a reported 4.6 PFLOPS of FP8 compute and 7.37 TB/s of bandwidth, per The Register and AI Compute Tracker. Maia 300 has no reported compute figure, and its 7 TB/s bandwidth claim, from TechTimes, is not officially confirmed.
Will Microsoft sell Maia 300 chips to other companies?
None of the current reporting indicates Microsoft plans to sell Maia 300 externally. Every report reviewed frames it as infrastructure for Microsoft’s own Azure and Copilot workloads, consistent with how Maia 200 is already used internally today, according to NerdLevelTech.
Why is TSMC packaging capacity such a big deal for Maia 300?
High-bandwidth-memory AI chips require advanced packaging that only a limited number of facilities can produce at volume. TechTimes reports Nvidia controls approximately 60% of TSMC’s advanced packaging capacity, meaning Microsoft has to compete for the remaining capacity alongside Google, Amazon, and every other AI chip designer, which directly constrains how fast Maia 300 can ramp toward the reported 300,000-unit 2027 target.
What happened to Microsoft’s original Maia chip production plan?
Reuters reported on June 27, 2025, citing The Information, that Microsoft’s next-generation Maia chip had its mass production delayed by at least six months, pushing what was originally a 2025 ramp into 2026. Maia 300 represents the recovery plan from that delay.
