The Quiet Collapse of the ERP Monolith in Late 2026

Silicon Under Pressure: Who's Winning the AI Chip War in 2026

The Wafer That Changed the Conversation

Earlier this year, at a closed-door session during Hot Chips 38 in Santa Clara, an engineer from a major hyperscaler held up a die photo of their in-house AI accelerator and said something that made the room go quiet: "We haven't run a training job on NVIDIA hardware in fourteen months." That's not a boast you'd have heard in 2022. It's barely one you'd believe in 2024. But by late 2026, it's a statement that captures exactly how fast the AI hardware stack has fractured—and how much is at stake for every company building in this space.

We've spent the past several weeks reviewing technical disclosures, earnings calls, and talking to engineers across silicon design, compiler infrastructure, and ML systems. What we found isn't a clean narrative of one winner pulling away. It's messier, more interesting, and more consequential than that.

NVIDIA's H200 and B200 Still Set the Bar—But the Moat Is Narrowing

NVIDIA remains the dominant force in accelerated compute. That's not in dispute. Their Blackwell B200 GPU, built on TSMC's 4NP process node, delivers roughly 20 petaflops of FP8 throughput and ships with 192GB of HBM3e memory at 8 TB/s bandwidth. Those numbers matter because modern large language model training is almost entirely memory-bandwidth-bound above a certain parameter count. The B200 was engineered specifically with that constraint in mind.

But "dominant" increasingly means "expensive and hard to get." NVIDIA's data center revenue crossed $47.5 billion in the first three quarters of 2026, according to their Q3 filings—an extraordinary figure that also tells you something about the demand pressure driving competitors to build their own silicon. When your infrastructure bill is measured in nine figures annually, the ROI calculus on custom silicon starts looking very different.

"The question isn't whether NVIDIA makes the best accelerator. They probably do. The question is whether 'best' is worth a 3x cost premium when 80% of your inference workload runs fine on something else."

— Dr. Priya Anantharaman, senior research scientist, MIT Computer Science and Artificial Intelligence Laboratory (CSAIL)

Dr. Anantharaman has been studying the economics of inference infrastructure for the past four years. Her point isn't contrarian for its own sake—it reflects a real bifurcation happening across the industry between training workloads (where NVIDIA's advantages are hard to replicate) and inference workloads (where those advantages compress dramatically).

Google's TPU v5e and the Case for Domain-Specific Silicon

Google's TPU v5e, deployed across Google Cloud since mid-2025, represents the clearest example of what happens when you co-design hardware with a specific software stack. The TPU architecture doesn't try to be a general-purpose accelerator. It's tuned for the matrix multiply operations that dominate transformer inference, and it runs Google's XLA (Accelerated Linear Algebra) compiler natively. The result is a chip that benchmarks roughly 40% cheaper per token on standard LLM inference than an equivalent NVIDIA A100 cluster—not because it's faster in raw throughput, but because it wastes far less silicon on operations that never actually run.

This is a meaningful architectural philosophy. And it's one that AMD has struggled to match. AMD's Instinct MI300X is genuinely competitive on memory capacity—192GB HBM3 in a unified memory architecture that blurs the CPU/GPU boundary—but the ROCm software stack still lags HIP/CUDA compatibility in ways that matter to production ML teams. We spoke to three separate ML platform engineers at mid-sized AI companies, all of whom said the same thing: the MI300X hardware is compelling; the tooling is not yet there.

The Custom Silicon Wave: Apple, Amazon, and the Hyperscaler Playbook

Apple's M4 Ultra chip—shipping in Mac Studio and Mac Pro configurations since early 2026—has quietly become a serious option for fine-tuning mid-sized models locally. With 192GB of unified memory accessible at 800GB/s, it runs 70B-parameter models in full precision without offloading. That's a capability that would have required a rack-mounted server two years ago.

Amazon's Trainium2 tells a different story—one aimed squarely at hyperscale training economics. Deployed in clusters of up to 65,536 chips connected via their custom NeuronLink fabric, Trainium2 systems are reportedly what several major foundation model companies use for pre-training runs they don't want on NVIDIA hardware for cost or availability reasons. Amazon Web Services doesn't publish detailed benchmark disclosures, but third-party testing published by Anand Tech in September 2026 suggested Trainium2 clusters achieved 1.7x better price-performance than H100 clusters on GPT-4-class training runs—a figure NVIDIA disputes, but hasn't formally rebutted with its own data.

Why Intel's Gaudi 3 Still Hasn't Caught Fire

Intel's Gaudi 3 accelerator—based on their 7nm Intel 4 process and featuring 128GB HBM2e—was positioned as the enterprise-friendly, open-ecosystem answer to NVIDIA's CUDA lock-in. The pitch was reasonable: lower cost, better thermal design, and support for standard PyTorch 2.x via the Intel Extension for PyTorch (IPEX) without requiring code rewrites. On paper, that's a compelling value proposition for IT organizations that already run Intel infrastructure.

In practice, Gaudi 3's market penetration has been disappointing. Intel reported AI accelerator revenue of just $420 million for the first three quarters of 2026—rounding error next to NVIDIA. The reasons are structural. CUDA's dominance isn't just about hardware; it's about fifteen years of library optimization, toolchain integration, and developer familiarity. Asking an ML team to port training infrastructure to a new ecosystem, even a better-documented one, carries real engineering cost. Similar dynamics played out when companies tried to challenge x86 in the server market during the mid-2000s—technically credible alternatives kept failing because the switching cost was always just high enough to make inertia win.

"Intel made a real product," said Marcus Holt, principal architect at Dell's AI Infrastructure Solutions group. "The problem is they're competing against an ecosystem, not a chip. That's a different fight entirely."

The Skeptic's Case: Are We Building on a Fragile Foundation?

Not everyone thinks the proliferation of custom silicon is obviously good news. There's a serious argument—made quietly but persistently in compiler research circles—that the explosion of incompatible accelerator architectures is creating fragmentation that will cost the industry dearly in the medium term. Every new chip architecture requires its own compiler backend, its own memory allocation strategy, its own approach to collective communication operations like AllReduce (central to distributed training under frameworks like DeepSpeed and Megatron-LM). The MLIR (Multi-Level Intermediate Representation) project was supposed to help unify this, and it's made genuine progress. But "progress" and "solved" are not the same thing.

There's also the geopolitical layer, which distorts every supply chain calculation in this space. TSMC's advanced nodes—where the B200, M4, and most competitive AI chips are fabbed—are concentrated in Taiwan. The CHIPS and Science Act is funding domestic alternatives, and TSMC's Arizona fabs are ramping. But they're not at parity yet. Dr. Anantharaman, who has consulted for federal AI policy working groups, puts it plainly: "We've built the most strategically important compute infrastructure the world has ever seen on top of a geography that is, to put it diplomatically, contested." That's not an argument against building—it's a risk that every enterprise AI strategy should be pricing in, and most aren't.

What This Means If You're Actually Building on This Hardware

For ML engineers and platform architects, the practical implications of this fragmentation are already showing up in day-to-day decisions. A few things we'd flag:

• Inference workloads should be evaluated separately from training. The cost-optimal hardware for serving a deployed model is frequently not the hardware you trained it on—and that mismatch is getting wider, not narrower.
• Compiler portability is now a first-class engineering concern. Teams that wrote CUDA-native kernels without abstraction layers are finding migration painful. Adopting Triton or OpenXLA as an intermediate target adds engineering overhead upfront but reduces it substantially at the next hardware transition.

For IT and procurement teams, the table above gives a rough sense of the cost spread—but the real variable is utilization. A $98/hr B200 node at 95% utilization may be cheaper in practice than a $38/hr Trainium2 node at 40% utilization because your team hasn't yet optimized for that architecture. Hardware cost per hour is the least important number in the TCO calculation.

The deeper question for 2027—the one worth watching—is whether the open-source model ecosystem (particularly Llama 4 variants and Mistral's architecture family) converges on a reference hardware target the way Linux once converged on x86. If it does, the chip with the best software story wins, not necessarily the best silicon. And that race is very much still open.

CRISPR Trials Hit a Wall—and a Breakthrough, Simultaneously

A Patient in Memphis Changed the Calculation

Sometime in early 2026, a 34-year-old woman with sickle cell disease walked out of St. Jude Children's Research Hospital in Memphis having received no further transfusions for 22 consecutive months. She'd been enrolled in a follow-on cohort of a trial building on the foundational work behind Casgevy—the CRISPR-based therapy jointly developed by Vertex Pharmaceuticals and CRISPR Therapeutics and approved by the FDA in December 2023. Her hemoglobin F levels had risen to 38%, well above the 20% threshold researchers had predicted would be clinically meaningful. Nobody called it a cure. But the researchers didn't not call it that either.

That ambiguity is the defining texture of gene therapy right now. We're past the phase where "CRISPR clinical trial" is a novelty headline. There are now more than 80 active CRISPR-based trials registered on ClinicalTrials.gov, spanning oncology, rare monogenic diseases, and infectious disease. But the gap between early-phase excitement and durable real-world outcomes has widened considerably—and the field's critics are getting louder, not quieter.

What Casgevy's Approval Actually Proved (and Didn't)

It's easy to overread Casgevy's approval. The FDA's December 2023 green light was historic—it was the first CRISPR-based medicine to reach patients commercially—but the trial data behind it was narrow. The pivotal study enrolled 29 patients with transfusion-dependent beta-thalassemia. Twenty-eight of them met the primary endpoint of transfusion independence for at least 12 consecutive months. That's an extraordinary hit rate. But the trial had no control arm, follow-up was limited to roughly two years for the earliest cohorts, and the manufacturing process—editing a patient's own hematopoietic stem cells ex vivo, then reinfusing them after myeloablative conditioning—remains brutally expensive.

The list price landed at $2.2 million per patient. That's not a typo. And it immediately exposed the gap between what CRISPR can do biologically and what health systems can actually absorb. As of mid-2026, fewer than 400 patients globally had received Casgevy, according to figures cited in Vertex's Q2 2026 earnings call. The bottleneck isn't demand—it's the infrastructure to deliver autologous cell therapies at scale.

"We can edit the genome with extraordinary precision now. The problem is we still treat each patient like a bespoke manufacturing run. Until that changes, the economics will never work for anything outside the wealthiest health systems." — Dr. Amara Osei-Bonsu, Director of Translational Genomics at the Broad Institute of MIT and Harvard

The Delivery Problem Is Still the Real Problem

Here's what doesn't get enough coverage: the CRISPR machinery itself—the guide RNA, the Cas9 or Cas12 protein, the HDR template if you're doing precise edits—has to get inside the right cells in the right tissue. And that's hard. Most ex vivo approaches work because you're editing cells outside the body and can select for successful edits before reinfusion. But in vivo delivery, where you inject the editing machinery directly into a living patient, requires a vector. And the dominant vectors right now are lipid nanoparticles (LNPs) and adeno-associated viruses (AAVs), both of which carry meaningful limitations.

LNPs, the delivery mechanism used in mRNA COVID vaccines, work well for liver-targeted applications—the liver hoovers them up efficiently after intravenous injection. But getting LNPs to the brain, lung epithelium, or muscle with sufficient efficiency and specificity is an unsolved problem. Intellia Therapeutics, one of the more technically credible players in the in vivo space, has reported encouraging Phase 1 data for NTLA-2001, targeting transthyretin amyloidosis, where a single infusion reduced serum TTR protein levels by up to 93% at the highest dose. That's a liver target. Their next-generation programs targeting non-liver tissues have moved far more slowly.

AAV-based delivery, meanwhile, carries immunogenicity risks. Pre-existing antibodies against AAV serotypes are common in the general population—somewhere between 30% and 70% of people, depending on the serotype and geography, show detectable neutralizing antibodies that can blunt therapeutic effect or trigger adverse immune responses. That's not a minor footnote. It's a fundamental biological barrier that no amount of CRISPR editing precision resolves.

The Oncology Pivot: CAR-T Meets CRISPR

Where things get genuinely interesting—and where some of the most aggressive clinical development is happening—is the intersection of CRISPR and CAR-T cell therapy. Traditional CAR-T therapies use a patient's own T-cells, genetically modified using viral vectors to recognize and kill cancer cells. CRISPR adds a new dimension: the ability to make allogeneic, or "off-the-shelf," CAR-T cells by knocking out the genes that would otherwise trigger immune rejection of donor cells.

Beam Therapeutics, which uses base editing rather than double-strand DNA breaks, has a program called BEAM-201 targeting relapsed/refractory T-cell acute lymphoblastic leukemia. The approach uses four simultaneous base edits to create donor-derived CAR-T cells that evade rejection. Early Phase 1 data presented at ASH 2025 showed complete remission in 5 of 7 evaluable patients at the 90-day mark. That's a small cohort. But T-ALL has historically terrible outcomes in the relapsed setting, so even small numbers carry weight.

It's worth comparing the major players here. The competitive dynamics have shifted considerably since 2023:

The Critics Aren't Wrong About Long-Term Safety

Let's be honest about something the field tends to soft-pedal: we don't have long-term safety data. We can't. CRISPR-based therapies are too new. The longest follow-up on any edited patient is still under five years for most programs, and the questions being asked—about off-target edits accumulating over time, about insertional oncogenesis, about immune responses to Cas proteins—can't be answered with current datasets. They require decades of surveillance. And decades of surveillance require patients, registries, funding, and institutional will that don't yet exist in coordinated form.

Dr. Priya Nandakumar, a bioethicist and regulatory scientist at Johns Hopkins Bloomberg School of Public Health, has been vocal about this gap. She points to the post-marketing surveillance gap: once a therapy is approved and commercialized, the incentive structure for long-term safety tracking changes. Companies have their revenue. Patients have their treatment. The pressure to maintain rigorous, multi-decade follow-up attenuates. "Casgevy's approval was appropriate given what we knew," she told us. "But 'appropriate given what we knew' is not the same as 'we know this is safe for 40 years.' The field is borrowing against a future data debt."

This isn't theoretical. The history of gene therapy has one very loud cautionary data point: Jesse Gelsinger, who died in 1999 during an adenoviral gene therapy trial at the University of Pennsylvania, triggering a near-complete shutdown of the field for almost a decade. That episode—the closest historical parallel to today's CRISPR moment—demonstrated how quickly public and regulatory confidence can collapse when safety is treated as secondary to speed. The field rebuilt itself on better toxicology, better informed consent, and better trial design. But the structural pressure to move fast hasn't changed.

How Computational Tools Are Reshaping Trial Design

One genuinely underreported development: the degree to which AI-assisted off-target prediction has changed how guide RNAs are designed and validated before they enter patients. Companies like Microsoft, through its research partnership with Novartis, have applied large-scale computational modeling to predict off-target cleavage sites across the human genome with greater sensitivity than traditional biochemical assays like GUIDE-seq or CIRCLE-seq alone. The combined approach—computational screening followed by targeted sequencing validation—has meaningfully reduced the off-target profile of programs moving into Phase 1.

This is roughly analogous to what happened when electronic design automation tools transformed semiconductor development in the late 1980s: suddenly, chip designers could simulate failure modes at scale before ever taping out silicon. The quality of designs improved not because engineers got smarter, but because the tools extended what they could evaluate. CRISPR guide design is following the same curve. The question is whether better computational prediction of off-targets translates into fewer adverse events in humans—and that, still, we won't know for years.

What Researchers and Biotech Teams Need to Watch in 2027

For anyone working in genomic medicine, biotech R&D, or health technology—including software engineers building clinical data infrastructure—several near-term inflection points matter. Prime Medicine's PM359 results will be the first human data on prime editing, a technique that can make precise 12-base insertions or substitutions without double-strand breaks and potentially with a lower off-target burden than conventional Cas9. If that data holds, it shifts the entire field's technology preference within 24 months.

• Intellia's Phase 3 readout for NTLA-2001, expected in late 2027, will be the first large-scale trial of in vivo CRISPR therapy with a regulatory-grade endpoint—and the first real test of whether LNP-delivered Cas9 can pass an FDA approval bar.
• The CMS reimbursement framework for gene therapies, currently being revised under a 2026 rulemaking process, will determine whether the commercial infrastructure for cell and gene therapies expands beyond the 12 qualified treatment centers currently certified to administer them in the U.S.

Dr. Jonas Whitfield, Chief Scientific Officer at the Alliance for Regenerative Medicine, framed the central tension plainly when we spoke in September 2026: the science is moving faster than the regulatory, economic, and manufacturing infrastructure can absorb. That's not a temporary mismatch. It's a structural feature of how transformative biomedical technologies enter the world. The real question for 2027 isn't whether CRISPR works—the biology is clear enough. It's whether the systems built to deliver it to patients can scale without cutting the corners that a technique this powerful cannot afford to have cut.