---
title: "Red Day for Team Red: AMD Drops 17% in Worst Session Since 2017 Despite Record Earnings"
description: AMD delivered the best quarter in its history and the market erased $30 billion in response, exposing the brutal arithmetic of being Nvidia’s only credible rival.
author: Darie Nani (Editor-in-Chief)
date: 2026-02-06T10:43:38.000Z
updated: 2026-02-26T18:01:31.592Z
canonical: https://www.sovereignmagazine.com/article/red-day-for-team-red-amd-drops-17-in-worst-session-since-2017-despite-record-earnings
image: https://cdn.nanimediahouse.com/jmwce3w7qkk.jpg
categories: Markets
content_type: Analysis
region: United States
publication: Sovereign Magazine
about:
  - type: Organization
    name: AMD
---

AMD delivered the best quarter in its 55-year history on 3 February 2026. Revenue hit $10.3 billion, beating Wall Street’s estimate by $600 million. Earnings per share came in at $1.53 against a forecast of $1.32. The stock dropped 17% the next morning, closing at $200.19, its worst single session since May 2017. Roughly $30 billion in market value disappeared between the opening bell and the close.

AMD is trapped in a market where being second to Nvidia means that record-breaking results are graded on a curve. Nvidia itself learned this in November, when it [beat Q3 estimates by $2.1 billion and still watched shares fall 3%](https://www.sovereignmagazine.com/article/why-nvidia-s-stellar-earnings-couldn-t-stop-the-tech-rout), triggering an Asian semiconductor rout that wiped billions from chip valuations globally. AMD’s punishment was five times worse for a similar offence: guidance that failed to match the acceleration investors had already priced in. In AI chips, steady growth at 32% year over year is apparently a disappointment.

## A one-time China windfall that hid the real run rate

Buried in the Q4 results was $390 million in Instinct MI308 chip sales to China. These shipments only happened because the US partially reversed its AI chip export restrictions, letting AMD clear a backlog of previously reserved inventory. The company also released $360 million in inventory that had been written down when those export licences were initially blocked.

Su said on the earnings call that the China revenue came from clearing old orders, not from new demand. First-quarter guidance includes just $100 million in continued MI308 China shipments (a 74% drop) with nothing further assumed. Susquehanna analyst Chris Rolland told CNBC that once you accounted for the China contribution, the Q4 beat was ‘far less substantial than we would have thought’.

Strip out the MI308 windfall and Q4 gross margins would have landed at roughly 55% rather than the reported 57%. That 55% figure is exactly what AMD guided for Q1, which means the underlying business did not deteriorate between quarters. It just was not accelerating at the rate the stock had priced in.

## The cost of being Nvidia’s only credible competitor

AMD is the only company shipping merchant silicon for AI training and inference at scale that is not Nvidia. Google, Amazon and Microsoft all build custom chips for their own data centres (and [OpenAI is now partnering with Broadcom to do the same](https://www.sovereignmagazine.com/article/the-ai-chip-wars-heat-up-how-openai-and-broadcom-are-reshaping-manufacturing-s-future)), but none of those are available for purchase by third parties. That makes AMD the default alternative, and the market is pricing it like the alternative should already be the equal.

Data centre segment revenue reached $5.4 billion in Q4, up 39% year over year. Lisa Su told CNBC that server CPU demand was ‘going gangbusters’ and that AI was ‘accelerating at a pace that I would not have imagined’. EPYC processors are gaining share against Intel across both hyperscaler and enterprise deployments.

But conventional measures do not apply when a stock has doubled in 12 months. Investors paying 40 times earnings need profit growth to outpace revenue growth, and AMD’s data centre operating margins have moved the wrong way (falling from 29% to 25% over recent quarters). The company says margins will reach 35% by late 2026. Four consecutive quarters of compression make that target harder to believe without evidence.

## Everything depends on the second half of 2026

AMD’s response to the margin pressure and growth questions is a product roadmap that concentrates almost all of its catalysts in the second half of this year. The Instinct MI400 GPU series, the Helios rack-scale server platform and the first phase of the OpenAI deployment all converge in that window.

The Helios system packs 72 GPUs into a single rack, delivering up to 1.4 exaFLOPS of FP8 performance with 31 TB of HBM4 memory. Those numbers put it in the same bracket as Nvidia’s Vera Rubin NVL72 system. The MI440X variant targets enterprise AI buyers who care more about cost per token and software compatibility than raw benchmark performance, a segment where AMD’s EPYC and Instinct lineups have already gained traction. Matt Kimball of Moor Insights and Strategy told Data Center Knowledge that enterprise AI was ‘the best long-term opportunity’ for AMD to differentiate from Nvidia’s hyperscaler dominance.

The largest single commitment comes from OpenAI. Under terms agreed in October 2025, OpenAI will deploy 6 gigawatts of AMD Instinct GPUs over multiple years, starting with a 1-gigawatt rollout in the second half of 2026. OpenAI may also take a 10% stake in AMD. Analysts estimate this first phase alone represents $15 billion to $20 billion in revenue, roughly half of AMD’s projected AI sales for 2026 and 2027 combined. That level of concentration on a single customer (one that is burning through $50 billion a year and seeking emergency funding) is a significant risk in itself.

## Supply chain bottlenecks that AMD cannot control

Even if demand materialises on schedule, AMD faces hardware constraints outside its control. The Helios rack requires UALink switches to move data between its 72 processors. Marvell, one of the primary switch suppliers, targets volume shipments for the second half of 2026 at the earliest. If those switches slip by six months, the entire OpenAI deployment timeline shifts with them.

Each MI450 chip requires 432 GB of HBM4 memory. SK Hynix, Samsung and Micron have sold their entire 2026 HBM production allocation, and Nvidia holds priority in the queue. The same [capacity constraints driving TSMC’s semiconductor dominance](https://www.sovereignmagazine.com/article/tsmc-s-ai-chip-leadership-drives-taiwan-s-semiconductor-dominance) apply across the AI chip supply chain: demand for advanced packaging and memory has outpaced fabrication capacity. SK Hynix’s chief financial officer said directly that the full 2026 supply was spoken for. AMD must compete for whatever allocation remains or wait for 2027 capacity to come online.

Su told analysts that the MI450 ramp was ‘on schedule to start in the second half of the year’ and that Helios was in ‘active discussions’ for additional sales. Nvidia has already responded: in January, it upgraded the memory bandwidth specifications for Vera Rubin, a pre-emptive move widely read as a direct counter to MI400’s performance claims. That AMD forced a spec revision from Nvidia before shipping a single MI400 unit suggests the competitive threat is real, but translating product parity into market share requires solving the memory and switch supply problems first.

## The wider signal for European procurement

European data centre capacity expanded 18% in 2025 according to CBRE, and procurement decisions over the next 12 months will lock in which silicon powers those facilities for years. AMD’s open ecosystem and competitive EPYC pricing have already won European cloud design wins, but the company’s ability to deliver rack-scale AI systems in volume is unproven.

A two-supplier market is better than a monopoly for European buyers negotiating AI infrastructure contracts. If AMD can ship Helios on time, European cloud operators gain leverage over Nvidia on pricing and supply terms. If AMD stumbles, Nvidia’s position strengthens at precisely the moment European governments are trying to build sovereign AI compute capacity.

Analyst consensus held at Moderate Buy after the sell-off, with an average price target of $281 (40% above the post-drop close). Cathie Wood’s ARK Invest bought the dip. Analysts raised EPS estimates for 2026 and 2027 by 3% and 9% respectively. The pattern is now consistent across the AI chip sector: Nvidia’s November sell-off, [Microsoft’s 6% drop after AI spending surged but Azure growth slowed](https://www.sovereignmagazine.com/article/microsoft-shares-drop-6-after-ai-spending-rises-and-cloud-growth-slows) and AMD’s February rout all followed strong results, all punished guidance that was merely good rather than exceptional, and all saw analysts hold or raise their targets within days. AMD must now prove, quarter by quarter, that the second half delivers what the roadmap promises. A company posting record revenue, record earnings and record data centre sales still lost 17% because guidance growth decelerated from Q4’s pace. That gap between execution and expectation defines what it costs to be the credible second player in AI chips.

## Further Context

**Q: Why did AMD stock fall 17% despite beating earnings estimates?**
AMD’s Q4 revenue of $10.3 billion exceeded Wall Street’s consensus by $600 million, and EPS of $1.53 beat the $1.32 forecast. The sell-off was driven by first-quarter 2026 guidance that showed a sequential revenue decline and gross margin compression from 57% to 55%. A one-time $390 million China chip shipment also inflated the Q4 beat, meaning underlying performance was closer to expectations than the headline figures suggested.

**Q: How much of AMD’s Q4 revenue came from China chip sales?**
AMD shipped $390 million in Instinct MI308 accelerators to China after the US partially reversed export restrictions, allowing the company to clear reserved inventory. First-quarter 2026 China revenue is expected to drop to approximately $100 million (a 74% sequential decline). AMD has not assumed additional China revenue beyond that amount.

**Q: How does AMD’s MI400 series compare to Nvidia’s Vera Rubin?**
AMD’s Helios rack system with MI400 GPUs claims up to 1.4 exaFLOPS of FP8 performance with 31 TB of HBM4 memory, placing it in a similar bracket to Nvidia’s Vera Rubin NVL72 racks. AMD’s strongest position is in enterprise AI where cost efficiency and software compatibility matter more than peak performance. Nvidia maintains advantages in hyperscaler relationships and priority access to HBM4 memory supply. Nvidia pre-emptively upgraded Vera Rubin’s specifications in January 2026 in response to AMD’s MI400 claims.

**Q: Will AMD recover from this decline?**
Analysts maintained a Moderate Buy consensus with a $281 average price target after the sell-off, roughly 40% above the post-drop close. Recovery depends on the MI400 GPU series and Helios rack-scale platform shipping on schedule in the second half of 2026, the OpenAI deployment beginning as contracted, UALink switches arriving on time and sufficient HBM4 memory becoming available.