Before diving into today’s brief… check out this week’s episode of The Compute 100 with Jeff Tatarchuk, Co-Founder and Chief Growth Officer at TensorWave!
Fresh off $350M in funding, we sit down with Jeff to discuss the bet on AMD, how AI has weakened CUDA’s moat, and the path to scaling the largest AMD clusters in the world. Thanks for joining us Jeff! Check out TensorWave for more details.
On to today’s stories!
New Epoch AI data latest data gives a preview of what is to come for the hyperscalers in the next quarter: the five largest hyperscalers — Microsoft, Amazon, Alphabet, Meta, and Oracle — will spend more on capital expenditure than they generate in operating cash flow in aggregate by Q3 2026. Capex is growing at roughly 70% annually against 23% for operating cash flow; Oracle has already crossed, Amazon is crossing now, and Alphabet, Meta, and Microsoft follow at intervals through Q3 2028. As the organic free-cash phase ends, external financing takes over — debt markets, equity raises, and cash drawdowns — shifting risk from hyperscaler balance sheets into the capital stack beneath them.
Atom Computing raised over $200 million in new capital on June 16 — a $100 million Series C led by Third Point Ventures (DCVC and Cisco Investments participating) and a $100 million Letter of Intent from the U.S. Department of Commerce. The company recently completed a full demonstration of quantum error correction on neutral-atom hardware, becoming only the second organization to do so and the first using neutral-atom technology; it is currently installing what it describes as the world's first commercial quantum computer with logical qubits, in partnership with Microsoft. The DOC commitment places fault-tolerant quantum infrastructure in US industrial policy's near-term build column.
The Department of Commerce signed a $50 million CHIPS Act LOI with Coherent on June 16 to expand its Sherman, Texas indium phosphide facility — doubling manufacturing space and quadrupling wafer capacity for the photonic devices that drive AI datacenter optical interconnects. The stock sold off despite the announcement as sector skepticism on co-packaged optics adoption timelines and pre-FOMC profit-taking dominated.
AMD signed a 30-megawatt definitive deployment agreement with Rackspace Technology on June 16 — AMD Instinct MI355X GPUs across Rackspace's global data centers for regulated enterprise workloads through 2028. The move is a named, contracted win for AMD's infrastructure stack; Rackspace (RXT) gained 12% on the same announcement while the broader sector sold off on FOMC positioning.
SemiAnalysis published the first results from its STEEL chip teardown lab this week: an analysis of Huawei's Kirin 9030, manufactured on SMIC's N+3 process. The finding — SMIC N+3's metal pitch is smaller than Intel 18A's, but transistor density lags Intel 18A by 38% — provides the first independent, hardware-measured data point on China's most advanced commercial node. For the AI chip supply chain, the practical read is that SMIC N+3 is not a near-term substitute for leading-edge TSMC or Samsung processes on density-constrained workloads, while remaining competitive on interconnect geometry in ways that matter for analog and connectivity applications.
June 16 saw the Department of Commerce sign CHIPS Act Letters of Intent extending the program beyond silicon: a $50 million LOI for Coherent's indium phosphide photonics facility and a $100 million LOI for Atom Computing's neutral-atom quantum infrastructure — alongside xLight's $150 million finalization for a free-electron laser EUV prototype earlier this month. Together, these reflect a policy judgment that AI infrastructure capability depends on enabling layers — new lithography sources, optical interconnects, and quantum coprocessors — not only on advanced silicon nodes.
FOMC rate decision at 2pm ET. Kevin Warsh holds his first press conference as Fed chair at 2:30pm ET. Markets price a hold at 3.50%–3.75%; the dot plot composition is the variable most likely to move compute-adjacent growth multiples.
Micron reports Q3 FY2026. Consensus: revenue $34.7 billion (vs. $6.8 billion a year ago), EPS $19.95 (vs. $1.98). The print is expected to set a record; Q4 guidance and HBM contract pricing updates are the signals to watch for AI memory supply trajectory.