AMD posts record training GPU revenue but faces margin compression from inference cost deflation and customer custom silicon adoption.
The disconnect between AMD's operational performance and its stock price has rarely been starker. When the chipmaker delivered second-quarter results on August 4, the market responded by knocking roughly 7–8% off the share price in after-hours trading. Revenue came in at $11.54 billion, up 50.1% year over year, with earnings per share of $1.66 topping analyst estimates and net income of $2.3 billion. Yet investors fixated on what fell short: the gross margin of 54% missed the roughly 56% analysts had penciled in, driven by the costly ramp-up of Helios infrastructure for AI applications. Capital expenditures also ballooned to $808 million, far exceeding earlier projections of $298 million. In a market that now dissects every decimal point of AI-related guidance, record sales alone no longer suffice.
The engine of AMD's growth remains unmistakable. Data center revenue surged more than 107% to $6.72 billion, powered by EPYC server processors and Instinct graphics chips for AI training. That segment now accounts for 58% of total company revenue, up sharply from a year earlier. CEO Lisa Su called the results excellent, while CFO Jean Hu signaled further acceleration in data center sales through the second half of 2026. For the current quarter, AMD guided to roughly $13 billion in revenue—a range of $12.7 billion to $13.3 billion, ahead of the $12.5 billion consensus. Su projected the server business should grow around 80% in the second half, with more than 70% growth expected in 2027 and a doubling of data center revenues.
Compounding post-earnings pressure was an external development unrelated to AMD's financials. Elon Musk declared that SpaceX would exclusively use Nvidia AI chips going forward, citing the competitor's "best architecture" and emphasizing a close partnership. Su responded diplomatically on CNBC, expressing "enormous respect" for Musk and pointing to a long-term partnership. The episode underscored how fragile investor confidence remains in the second force of the AI chip market—a single public endorsement of the competition can outweigh a strong quarter.
Two days after earnings, AMD announced the acquisition of Taalas, a Toronto-based startup specializing in model-specific chips for AI inference. Taalas's core product, the HC1 chip, bakes model weights directly into the chip architecture, eliminating the need for external memory. For Llama 3.1 with 8 billion parameters, the chip reportedly delivers roughly 17,000 tokens per second per user with significantly lower energy consumption than conventional GPU solutions. The technology will feed into AMD's Instinct graphics processors, bolstering the company's inference strategy against Nvidia—which had earlier acquired the startup Groq.
Analyst reactions painted a picture of genuine disagreement. Rosenblatt Securities raised its price target to $700 from $665, while Argus Research jumped to $625 from $450. Jefferies reaffirmed a $650 target, JPMorgan sees fair value at $550, and Wells Fargo remains most bullish at $700. The consensus in early August showed 36 buy ratings, five "outperform" calls, and ten holds—with zero sell ratings. The median price target stood near $579, roughly 22% above the prevailing share price, with individual targets reaching as high as $1,250. Not all were convinced: a Seeking Alpha analyst issued a Sell rating, arguing that the gross margin fell weaker than expected and that current valuation already prices in near-perfect execution of the growth strategy.
Institutions hold roughly 71.34% of outstanding shares. CEO Lisa Su disposed of 125,000 shares on June 10 at $460.69 each. Insider sales over the past 90 days totaled approximately $130.4 million—routine after strong rallies rather than a signal of lost confidence.
By Friday's close in German trading, AMD shares sat at €418.00, down 1.54% on the day. Over the past 30 days, the stock has shed 7.70% and trades 6.83% below its 50-day moving average, remaining 18.31% off its 52-week high. The pullback looks less like a fundamental break and more like a technical pause following an extraordinary twelve-month rally. The next earnings report arrives November 3. Between now and then, the market will watch whether Hu's promised acceleration in data center sales materializes and whether Taalas integration starts showing tangible progress in the race for AI inference chips.