AMD Instinct Analysis: Can MI350 Become the #2 AI Accelerator Platform?
AMD: Can Instinct Become the #2 AI Accelerator Platform?
Can AMD turn Instinct from a credible Nvidia alternative into the industry's clear #2 AI accelerator platform, or will CUDA and Nvidia's full-stack advantage remain too difficult to overcome?
Executive Summary
AMD's data center business has gone from a rounding error next to Nvidia to a $6.7 billion-a-quarter operation in roughly two years, and the second quarter of 2026 made the shift impossible to ignore: data center revenue more than doubled year-over-year, driven by AMD's EPYC server CPUs and Instinct MI350-series GPUs. Behind that number sits a run of hyperscaler commitments that would have been unthinkable for AMD's accelerator business a year ago — OpenAI, Meta, Microsoft, Oracle, and Anthropic have all signed multi-gigawatt supply agreements, some structured with equity warrants tied directly to shipment milestones.
And yet the same earnings print that produced those numbers also triggered a sharp stock selloff, not because of anything AMD did wrong, but because Elon Musk told SpaceX investors, hours later, that his company would build its AI infrastructure "exclusively" on Nvidia. That contrast — record hyperscaler wins on one hand, a single competitive comment moving the stock on the other — captures where AMD actually stands: a legitimate, well-capitalized second source with real design wins, still operating in Nvidia's shadow on both revenue scale and software maturity.
The honest read is that AMD has solved the hardware credibility problem. The MI350 series matches or beats Nvidia's Blackwell B200 on paper in memory capacity and competes on inference throughput. What AMD has not yet solved is the software and ecosystem gap that CUDA has spent nineteen years building — and that gap, more than any single chip generation, is still the thing standing between Instinct and a clear #2 position.
AMD AI Strategy
AMD's Instinct roadmap has moved from an annual cadence to something closer to Nvidia's own yearly refresh rhythm, a deliberate choice to stop giving Nvidia multi-year windows between generations. The MI300 series established AMD as a credible inference alternative in 2023-2024. The MI350 series, launched in mid-2025 and now the primary revenue driver in AMD's data center segment, pushed memory capacity and bandwidth ahead of Nvidia's Blackwell B200 on paper. The next step, the MI450 series built on CDNA 5 and TSMC's 2nm process, was unveiled at AMD's Advancing AI 2026 event alongside the Helios rack-scale system — and it is the MI450, not the MI350, that anchors AMD's largest hyperscaler commitments.
The strategic logic is to stop competing chip-for-chip and start competing rack-for-rack. AMD's Helios platform bundles Instinct GPUs, EPYC "Venice" CPUs, and Pensando "Vulcano" networking into a single reference architecture, mirroring Nvidia's own move from selling individual GPUs to selling entire NVL72-class systems. For AMD to become a clear #2, three things need to happen simultaneously: MI450-class hardware needs to hold its performance claims under independent benchmarking, ROCm needs to close enough of the software gap that switching from CUDA is not a multi-quarter engineering project, and the gigawatt-scale hyperscaler commitments already signed need to convert into recurring, high-margin revenue rather than one-off capacity deals.
MI350 Analysis
The Instinct MI350 series — the MI350X air-cooled part and the MI355X liquid-cooled flagship — is built on AMD's fourth-generation CDNA architecture on a 3nm process, packing 185 billion transistors. Both variants ship with 288GB of HBM3E memory and 8TB/s of bandwidth, ahead of the 192GB on Nvidia's Blackwell B200 (Nvidia's GB200 superchip pairs two Blackwell dies for 384GB, narrowing that gap at the system level). The MI355X draws up to 1,400 watts, matching Blackwell Ultra's power envelope — a sign that the two platforms have converged on similar infrastructure requirements even where the underlying silicon differs.
AMD's own benchmarks claim up to 4x AI compute performance and a 35x inference improvement over the prior MI300X generation (that inference figure applies specifically to the MI355X against MI300X, not against any Nvidia part). Independent results are more measured but still credible: at MLPerf Inference 6.0, published in April 2026, MI355X posted results within single-digit percentage points of B200 on server inference workloads — a meaningful outcome because MLPerf uses standardized submission rules across vendors, making it one of the few apples-to-apples comparisons available. On raw FP8 throughput, MI350X roughly matches B200; on memory capacity, it exceeds it, letting a single MI350X fit larger models that would otherwise require multiple B200s and the multi-GPU coordination overhead that comes with them.
Where AMD still trails is model FLOPS utilization — the share of theoretical peak performance a system actually delivers in production. Estimates put Nvidia's H100/B200 generation at 50-55% MFU at scale, versus roughly 45% for AMD's MI300X-class hardware, a gap driven less by silicon than by years of CUDA kernel optimization and NCCL's maturity for multi-node scaling. That is the crux of the MI350 story: the chip-level spec sheet increasingly favors AMD on memory, and inference benchmarks are closing, but system-level, production-scale performance still shows a real if narrowing NVIDIA advantage — which is precisely why AMD's pitch has shifted from individual chip specs to full-rack Helios systems.
On availability, MI350 shipped to partners and hyperscale data centers beginning in Q3 2025 and became AMD's primary data-center revenue driver through the first half of 2026, alongside continuing MI300X and MI325X volume. AMD also introduced the MI350P, a PCIe-based part aimed at mainstream enterprise deployment rather than large training clusters, extending Instinct's addressable market below the hyperscaler tier.
ROCm & Software Ecosystem
Software remains AMD's most honest competitive disadvantage, and AMD's own actions confirm it: the company has moved ROCm to an accelerated, six-week release cadence — eight to nine releases a year, versus the roughly annual cycle it ran historically — specifically to close feature and performance parity gaps with CUDA faster than a single yearly update ever could. The July 2026 release, ROCm 7.14, marked the production debut of "TheRock," a new automated, open-source build and release system meant to let AMD ship official hardware support in weeks rather than an entire quarter after new silicon launches. ROCm 7 also added native Windows and WSL2 support for the first time, a belated but strategically important move to reach the much larger population of AI developers who work primarily in Windows environments rather than dual-booting Linux.
Independent analysis still quantifies a real gap. One widely cited benchmark, the "CUDA Gap Score," estimates that CUDA's software optimization alone is worth the equivalent of 30-99% more effective hardware performance across various real-world workloads — a reminder that ROCm's job is not just matching CUDA's raw feature list but matching nineteen years of kernel-level tuning, framework integration, and a developer base numbering in the millions. AMD's HIP layer (Heterogeneous-compute Interface for Portability) is designed to lower the switching cost by letting CUDA-like code port to AMD GPUs with minimal changes, and PyTorch now has first-class ROCm support, which matters because PyTorch is the dominant framework for both research and production AI development.
The clearest external validation of ROCm's progress is not a benchmark — it's Meta's willingness to commit 6 gigawatts of production capacity to AMD silicon. Meta's ML infrastructure organization is not known for taking casual software-stack risk at that scale; a commitment of that size implies Meta's own internal testing found ROCm reliable enough for production training and inference workloads, not just experimental deployment. That said, "reliable enough for a sophisticated hyperscaler with a dedicated optimization team" is a different bar than "ready for a mid-market enterprise team with no in-house GPU specialists" — and the latter is where CUDA's ecosystem advantage remains largely intact.
Hyperscaler Adoption
The past twelve months have produced the most significant hyperscaler commitments in AMD's history, and it's worth separating the announced scale from what has actually shipped:
- OpenAI — committed up to 6 gigawatts of AMD GPU capacity under a multi-generation agreement signed in October 2025, with the first gigawatt of MI450 deployments beginning in the second half of 2026. The deal includes a performance-based warrant for AMD shares tied to shipment and stock-price milestones, a structure AMD has since replicated with other partners.
- Meta — signed a matching up-to-6-gigawatt agreement in February 2026, centered on a custom MI450-based GPU co-engineered specifically for Meta's inference workloads, paired with EPYC "Venice" CPUs on the Helios rack architecture. The deal carries a potential value exceeding $100 billion and includes a warrant for 160 million AMD shares — roughly a 10% stake if fully vested. Notably, Meta signed a separate large-scale Blackwell and Vera Rubin agreement with Nvidia just a week earlier, making Meta the clearest example yet of a hyperscaler running a genuine dual-vendor GPU strategy rather than picking a single supplier.
- Oracle — committed to deploying 50,000 MI450-series GPUs beginning in Q3 2026, positioning Oracle Cloud Infrastructure as the first hyperscaler to offer a publicly available AI supercluster built on AMD's next-generation platform, with further expansion planned into 2027.
- Microsoft — expanded its collaboration with AMD to deploy Helios racks and sixth-generation EPYC CPUs at scale across Azure, announced alongside AMD's Q2 2026 results.
- Anthropic — announced a strategic agreement in the same quarter to deploy up to 2 gigawatts of MI450-series GPUs in Helios racks, adding a fifth major AI lab or hyperscaler to AMD's committed pipeline.
Taken together, these agreements put roughly 20 gigawatts of announced AMD accelerator capacity on the board across five of the industry's most important AI buyers — an extraordinary figure (a single gigawatt is roughly the output of a nuclear power plant). AMD has said each gigawatt of deployed capacity should generate "significant double-digit billions" in revenue, which is the basis for the company's increasingly aggressive long-term guidance.
The distinction that matters for judging whether this is real demand or announcement inflation: none of these are pilot programs. They are structured as multi-year, multi-generation supply agreements with shipment milestones, several tied to equity compensation that only vests if AMD actually delivers hardware on schedule — a structure that gives both sides real financial skin in the execution, not just a press-release relationship. The near-term test is whether the "first gigawatt" deployments landing in the second half of 2026 for OpenAI, Meta, and Oracle actually ship on the disclosed timelines and perform at the promised level once independent users get their hands on them.
NVIDIA Competitive Moat
Nvidia's advantage is not a single moat but a stack of them, and they are not equally vulnerable to AMD's current strategy:
- CUDA and software: Nineteen years of kernel optimization, a massive trained developer base, and default-path integration into every major AI framework. This is the moat AMD is most directly attacking with ROCm's accelerated release cadence and HIP portability layer — and the one least likely to close quickly, because it compounds with every new model architecture Nvidia optimizes for first.
- Networking: NVLink and NCCL give Nvidia tightly integrated, high-bandwidth multi-GPU scaling that AMD's Pensando-based networking is still catching up to at the largest cluster sizes — a meaningful factor in large-scale training specifically, less so in single-node or small-cluster inference.
- Systems: Nvidia's shift from selling GPUs to selling full NVL72-class racks (and now Vera Rubin-generation systems) raised the competitive bar to full-stack integration — which is exactly the level AMD is now trying to match with Helios.
- Developer adoption and enterprise relationships: Nvidia's relationships span from individual researchers training models on a single GPU to the largest hyperscalers running frontier training runs — a breadth advantage AMD is only beginning to replicate below the hyperscaler tier.
- Scale and margin: Nvidia's data center revenue reached roughly $75 billion in a single quarter in early 2026 (fiscal Q1 2027), against AMD's data center segment revenue of $6.7 billion for the same period — a gap of roughly 11x that funds Nvidia's ability to out-invest AMD in R&D, supply chain commitments, and go-to-market across every layer of the stack simultaneously.
Of these, AMD has the most realistic near-term path to eroding Nvidia's advantage on systems (Helios is a credible full-rack answer to NVL72) and on price/memory-per-dollar for inference workloads, where MI350's larger HBM pool gives it a genuine cost advantage on models that would otherwise require more GPUs on Nvidia hardware. CUDA's software moat and Nvidia's sheer scale advantage are the two elements least likely to close within this product cycle.
AMD's Full-Stack AI Position
AMD's pitch to hyperscalers increasingly rests on being the only company besides Nvidia that can offer CPUs, GPUs, networking, and software as a single, co-engineered platform rather than components a customer has to integrate themselves. EPYC server CPUs (soon the sixth-generation "Venice" line) pair with Instinct GPUs and Pensando "Vulcano" networking inside the Helios rack architecture — the same bundling logic behind Nvidia's Grace-Blackwell systems. AMD's data-center CPU business is not a side note here: EPYC's continued strength was one of the two named drivers (alongside Instinct) of the 107% year-over-year data center revenue growth in Q2 2026, and it gives AMD a second, higher-margin, less-contested revenue stream to fund GPU R&D and cross-sell into accounts that might otherwise buy Intel Xeon and Nvidia GPUs from two separate vendors.
That combination — credible CPUs plus improving GPUs plus a rack-scale reference design plus a co-development relationship on the software side — is arguably AMD's strongest argument for being a full-stack alternative rather than a discount GPU supplier. It is also, notably, the argument every one of AMD's recent hyperscaler deals validates: Meta, Oracle, and Anthropic are all deploying EPYC and Instinct together inside Helios racks, not buying Instinct GPUs to pair with someone else's CPUs.
Competitive Landscape
NVIDIA
Still the default choice by a wide margin. Independent market-share estimates put Nvidia at roughly 80-87% of the AI accelerator market by revenue against AMD's 5-13%, depending on methodology and whether custom hyperscaler silicon is counted separately. Nvidia's Q1 FY2027 data center revenue of roughly $75 billion dwarfs AMD's $5.8-6.7 billion quarterly data center runs across the same period. Nvidia's edge is deepest in large-scale training and in any workload where model architectures are still evolving quickly enough that CUDA's first-mover optimization matters; AMD's edge is narrowest exactly there and widest in memory-bound inference.
Broadcom
Broadcom is not a direct accelerator competitor to AMD in the way Nvidia is — it's the dominant design partner behind hyperscaler-built custom ASICs (Google's TPU, Meta's MTIA, Microsoft's Maia, and reported programs for OpenAI and Anthropic), commanding an estimated 60-70% share of that design-services market alongside Marvell. That distinction matters strategically: Broadcom's threat to AMD is indirect but real — every dollar a hyperscaler spends on a Broadcom-designed custom chip for its own internal inference workloads is a dollar that never enters the merchant-GPU market AMD and Nvidia both compete for. Custom ASIC-based AI server shipments are projected to reach roughly 28% of the total AI server market in 2026, growing nearly three times faster than merchant GPU shipments — a structural headwind for AMD's total addressable market that has nothing to do with AMD's own execution.
Intel
Intel's Gaudi accelerator line has won selective Microsoft Azure and government contracts but has not displaced custom hyperscaler silicon or meaningfully dented AMD or Nvidia's share at scale. Intel's more significant long-term relevance to this competitive picture may come through its foundry business rather than its accelerators — if Intel Foundry Services can offer a credible, lower-cost alternative to TSMC's advanced nodes, it could reshape the cost structure for AMD, Nvidia, and Broadcom's custom-silicon customers alike, but that remains a multi-year bet rather than a near-term competitive factor.
Hyperscaler custom silicon
This is arguably the more structurally important long-term competitor to AMD, not because it competes for the same deals, but because it removes demand from the market AMD is trying to win. Google's TPU v7 ("Ironwood"), Amazon's Trainium 3, Microsoft's Maia, and Meta's MTIA are all now in production or scaling, each optimized narrowly for its own hyperscaler's workloads and reportedly delivering 30-50% lower total cost of ownership than merchant GPUs for those specific, high-volume, predictable use cases. Custom silicon's economics work best precisely for the largest, most predictable inference workloads — the same segment where AMD's memory-capacity advantage would otherwise be most competitive, creating direct overlap between AMD's best opportunity and hyperscalers' strongest incentive to build their own chips instead of buying anyone's.
Growth Drivers
- MI450 ramp: The first gigawatt-scale deployments for OpenAI, Meta, and Oracle land in the second half of 2026 — the single largest near-term catalyst for AMD's data center revenue trajectory.
- HBM4 availability: MI450's 432GB of HBM4 at 20TB/s bandwidth is a genuine memory-capacity leap that lets customers train and infer models roughly 50% larger entirely in-memory — a direct lever on AMD's inference cost advantage if supply keeps pace with demand.
- Hyperscaler deployment conversion: Turning the roughly 20 gigawatts of announced commitments into recurring, on-schedule shipments is the single biggest value-unlock available to AMD over the next 18 months.
- ROCm's accelerated cadence: Eight to nine releases a year, plus native Windows/WSL2 support and the new TheRock build system, meaningfully shortens the time between new hardware launches and full software support — narrowing, even if not closing, the CUDA gap.
- Inference-led AI demand: As industry compute shifts from training toward serving models at scale, AMD's memory-capacity advantage and "tokens-per-dollar" positioning matter more, since inference is less dependent on the multi-node training-scale software optimization where CUDA's edge is largest.
- EPYC + Instinct cross-sell: Every Helios rack deployment sells EPYC CPUs alongside Instinct GPUs, giving AMD a second growth engine and a stronger full-stack pitch against Nvidia's Grace-Blackwell bundle.
Risks
- Nvidia's continued product leadership: Blackwell Ultra and the newly announced Vera Rubin platform mean AMD is chasing a moving target, not a fixed one — every MI-series launch has to beat a Nvidia generation that has also moved forward.
- CUDA lock-in and the SpaceX moment: Elon Musk's on-the-record comment that SpaceX will build "exclusively" on Nvidia because it offers "the best architecture" — made hours after AMD's own record earnings report — is a pointed reminder that even sophisticated, technically elite AI buyers can and do choose single-vendor CUDA lock-in over multi-vendor diversification when they believe the software and system advantage is decisive.
- ROCm adoption below the hyperscaler tier: Meta and OpenAI have the in-house engineering depth to make ROCm work at scale; the much larger population of mid-market enterprises and smaller AI teams does not, and that's where CUDA's ecosystem advantage remains largely uncontested.
- HBM and advanced packaging constraints: TSMC's 3nm capacity is reportedly running at 100% utilization with demand roughly three times available supply, and HBM manufacturers face similar constraints — a shared industry bottleneck that could delay AMD's MI450 ramp regardless of demand or execution.
- Customer concentration in committed capacity: A large share of AMD's forward revenue visibility now sits in a handful of multi-gigawatt agreements with OpenAI, Meta, Oracle, Microsoft, and Anthropic — a structure that provides unusual revenue visibility but also means any one partner slowing its buildout, as SpaceX effectively signaled it would favor Nvidia instead, has an outsized effect on sentiment and, potentially, on realized revenue.
- Hyperscaler custom silicon: Broadcom- and Marvell-enabled ASIC programs at Google, Amazon, Microsoft, and Meta are growing nearly three times faster than the merchant GPU market and compete most directly in the large-scale, predictable inference workloads where AMD's memory advantage would otherwise be strongest.
- AI infrastructure spending normalization: Combined hyperscaler capex of $660-690 billion in 2026 is an extraordinary run rate; any broad-based pullback in AI infrastructure spending would compress demand for AMD and Nvidia alike, with AMD's newer, less diversified customer base potentially more exposed than Nvidia's.
- Rapid technology cycles: AMD has moved to a near-annual Instinct cadence specifically to keep pace with Nvidia, which raises execution risk — a delayed or underperforming MI450 launch would be far more damaging now that so much forward revenue is tied to its specific, disclosed shipment timelines.
- Valuation risk: AMD trades at a significantly higher forward P/E than Nvidia (high-50s to 80s-plus, depending on the period, versus Nvidia's 25-33x), pricing in continued execution on a roadmap that is still substantially unproven at scale.
Strategic Positioning
AMD's realistic strategic position is "credible, well-capitalized second source" rather than "emerging co-leader," and the company's own hyperscaler deals reflect that framing more than AMD's marketing does. Meta signed its AMD deal one week after signing a larger Nvidia deal. Microsoft and Oracle are deploying AMD alongside, not instead of, continued Nvidia purchases. That is precisely the role hyperscalers want AMD to play: a second supplier credible enough to extract better pricing and reduce single-vendor risk from Nvidia, without yet being asked to be anyone's primary AI infrastructure vendor.
The more interesting strategic question is whether that role is a ceiling or a stepping stone. AMD's hardware roadmap has genuinely closed the memory-capacity and inference-throughput gap; its rack-scale Helios platform is a credible full-stack answer to Nvidia's own systems strategy; and its accelerated ROCm cadence shows real urgency about the software problem rather than complacency about hardware alone being enough. What AMD has not yet proven is that ROCm can extend Instinct's appeal beyond the small number of hyperscalers with the engineering depth to make any GPU platform work — and until that happens, "#2 AI accelerator platform" describes AMD's position among the five or six companies capable of deploying compute at gigawatt scale, not its position across the broader enterprise AI market where CUDA's ease-of-use advantage remains largely untested by AMD's current strategy.
Key Financial & Operating Metrics (as of Q2 2026, reported August 4, 2026)
| Metric | Value |
|---|---|
| Q2 2026 total revenue | $11.5 billion, up 50% YoY (company record) |
| Data Center segment revenue | $6.7 billion, up 107% YoY (58% of total revenue) |
| Data Center operating income | $2.1 billion (vs. an operating loss a year earlier) |
| Non-GAAP EPS | $1.66 (vs. ~$1.61 consensus) |
| Non-GAAP gross margin | 56% (up from 55% in Q1 2026) |
| Q3 2026 revenue guidance | ~$13 billion (~41% YoY) |
| MI350 series memory / bandwidth | 288GB HBM3E, 8TB/s (vs. Nvidia B200: 192GB) |
| MI450 series memory / bandwidth | 432GB HBM4, 20TB/s |
| Announced hyperscaler commitments | ~20GW combined across OpenAI (6GW), Meta (6GW), Oracle (50,000 MI450 GPUs), Anthropic (2GW), plus Microsoft Azure expansion |
| AI accelerator market share (est.) | AMD ~5-13% vs. Nvidia ~80-87%, depending on methodology |
| Market cap (as of August 7, 2026) | ~$789-798 billion |
| Stock reaction to Q2 2026 print | Fell ~7-10% despite the earnings beat, largely attributed to Elon Musk's same-day comment that SpaceX would build exclusively on Nvidia |
| 12-month stock performance (as of Aug. 2026) | Up roughly 130-200%, outpacing Nvidia's ~10-20% gain over the same period |
Company Analysis Conclusion
AMD has done the hard part that most analysts thought would take much longer: it built accelerators that compete with Nvidia on paper, won public, structurally serious commitments from five of the industry's most important AI buyers, and forced Nvidia into a genuine two-vendor conversation at companies like Meta that would have been unthinkable eighteen months ago. That is a real achievement, not a marketing narrative — the gigawatt-scale, warrant-backed deal structures with OpenAI and Meta put financial teeth behind AMD's execution in a way that ordinary purchase orders would not.
But "credible #2 supplier to a handful of hyperscalers with deep in-house engineering" and "the industry's clear #2 AI accelerator platform" are not the same claim, and the gap between them is exactly the CUDA and ecosystem advantage this analysis keeps returning to. ROCm's accelerated release cadence and the Meta deal both suggest AMD is closing that gap faster than skeptics expected. Elon Musk's same-day rebuke — choosing Nvidia specifically for its architecture, not its availability — is a reminder that the gap has not closed yet, and that sophisticated buyers with a real choice are still choosing Nvidia when software and system maturity are what matters most. AMD's path to the clear #2 position runs directly through making that choice look wrong more often over the next several product cycles — not through winning another headline hyperscaler deal, which AMD has already shown it can do.
Source Attribution
- AMD Newsroom — AMD Reports Second Quarter 2026 Earnings
- SEC EDGAR — AMD Form 10-Q, Q2 2026
- TechPowerUp — AMD Reports Second Quarter 2026 Financial Results
- CNBC — Lisa Su Brushes Off Musk's Nvidia Commitment as AMD Stock Sinks After Earnings
- TheStreet — AMD Stock Falls After Record Quarter as SpaceX Picks Nvidia
- Introl Blog — AMD MI350 GPU Competition
- Spheron Network — AMD MI350X vs NVIDIA B200: Specs, Benchmarks, and Cloud Pricing
- SiliconAnalysts — AMD vs NVIDIA AI GPU Market Share 2026
- AMD — AMD Instinct MI350 Series GPUs
- AMD ROCm Blog — ROCm 7.14: TheRock Goes Production
- AIMultiple — GPU Software for AI: CUDA vs. ROCm in 2026
- Spheron Network — ROCm vs CUDA: AMD vs NVIDIA AI Stack Compared
- WION — AMD Unveils MI450 Superchip, Gets OpenAI's 6-Gigawatt Backing
- IntuitionLabs — OpenAI-AMD AI Hardware Partnership: A Strategic Analysis
- Introl Blog — Meta's $100B AMD Deal
- Jon Peddie Research — AMD and Meta Announce $60B, 6GW AI Partnership
- Oracle — Oracle and AMD Expand Partnership for Next-Generation AI Scale
- Tom's Hardware — The Custom AI ASIC State of Play (May 2026)
- Hashrate Index — The Two Companies Behind Hyperscaler AI Silicon
- The Brand Hopper — Broadcom's Top Competitors & Rivals
- 247wallst — How AMD Captures Significant Market Share From Nvidia
- companiesmarketcap.com — AMD Market Capitalization
Figures reflect publicly reported data as of August 9, 2026. Market share, stock price, and market-cap figures move frequently and were sourced from analyst estimates that vary by methodology — verify current values before publishing if this piece runs more than a few days after drafting.
Reviewed by Erwin Castro
on
Sunday, August 09, 2026
Rating: